Upload
zukun
View
475
Download
3
Tags:
Embed Size (px)
Citation preview
Visual Recognition in Primates and Machines
Tomaso
Poggio (with Thomas Serre)McGovern Institute for Brain Research
Center for Biological and Computational LearningDepartment of Brain amp Cognitive Sciences
Massachusetts Institute of TechnologyCambridge MA 02139 USA
NIPSrsquo07 tutorial
(preliminary)
Motivation for studying vision trying to understand how the brain works
bull
Old dream of all philosophers and more recently of AI ndash
understand how the brain works
ndash
make intelligent machines
This tutorial using a class of models to summarizeinterpret
experimental results
bull
Models are cartoons of reality eg
Bohrrsquos model of the hydrogen atom
bull
All models are ldquowrongrdquobull
Some models can be useful summaries of data and some can be a good starting point for more complete theories
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
The problem recognition in natural images (eg ldquois there an animal in the imagerdquo)
How does visual cortex solve this problem How can computers solve this problem
Desimone amp Ungerleider
1989
dorsal streamldquowhererdquo
ventral streamldquowhatrdquo
A ldquofeedforwardrdquo version of the problem rapid categorization
Movie courtesy of Jim DiCarlo
Biederman
1972 Potter 1975 Thorpe et al 1996
SHOW RSVP MOVIE
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
Modified from (Gross 1998)
A model of the ventral stream which is also an algorithm
[software available online]
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Motivation for studying vision trying to understand how the brain works
bull
Old dream of all philosophers and more recently of AI ndash
understand how the brain works
ndash
make intelligent machines
This tutorial using a class of models to summarizeinterpret
experimental results
bull
Models are cartoons of reality eg
Bohrrsquos model of the hydrogen atom
bull
All models are ldquowrongrdquobull
Some models can be useful summaries of data and some can be a good starting point for more complete theories
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
The problem recognition in natural images (eg ldquois there an animal in the imagerdquo)
How does visual cortex solve this problem How can computers solve this problem
Desimone amp Ungerleider
1989
dorsal streamldquowhererdquo
ventral streamldquowhatrdquo
A ldquofeedforwardrdquo version of the problem rapid categorization
Movie courtesy of Jim DiCarlo
Biederman
1972 Potter 1975 Thorpe et al 1996
SHOW RSVP MOVIE
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
Modified from (Gross 1998)
A model of the ventral stream which is also an algorithm
[software available online]
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
This tutorial using a class of models to summarizeinterpret
experimental results
bull
Models are cartoons of reality eg
Bohrrsquos model of the hydrogen atom
bull
All models are ldquowrongrdquobull
Some models can be useful summaries of data and some can be a good starting point for more complete theories
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
The problem recognition in natural images (eg ldquois there an animal in the imagerdquo)
How does visual cortex solve this problem How can computers solve this problem
Desimone amp Ungerleider
1989
dorsal streamldquowhererdquo
ventral streamldquowhatrdquo
A ldquofeedforwardrdquo version of the problem rapid categorization
Movie courtesy of Jim DiCarlo
Biederman
1972 Potter 1975 Thorpe et al 1996
SHOW RSVP MOVIE
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
Modified from (Gross 1998)
A model of the ventral stream which is also an algorithm
[software available online]
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
The problem recognition in natural images (eg ldquois there an animal in the imagerdquo)
How does visual cortex solve this problem How can computers solve this problem
Desimone amp Ungerleider
1989
dorsal streamldquowhererdquo
ventral streamldquowhatrdquo
A ldquofeedforwardrdquo version of the problem rapid categorization
Movie courtesy of Jim DiCarlo
Biederman
1972 Potter 1975 Thorpe et al 1996
SHOW RSVP MOVIE
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
Modified from (Gross 1998)
A model of the ventral stream which is also an algorithm
[software available online]
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
The problem recognition in natural images (eg ldquois there an animal in the imagerdquo)
How does visual cortex solve this problem How can computers solve this problem
Desimone amp Ungerleider
1989
dorsal streamldquowhererdquo
ventral streamldquowhatrdquo
A ldquofeedforwardrdquo version of the problem rapid categorization
Movie courtesy of Jim DiCarlo
Biederman
1972 Potter 1975 Thorpe et al 1996
SHOW RSVP MOVIE
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
Modified from (Gross 1998)
A model of the ventral stream which is also an algorithm
[software available online]
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
How does visual cortex solve this problem How can computers solve this problem
Desimone amp Ungerleider
1989
dorsal streamldquowhererdquo
ventral streamldquowhatrdquo
A ldquofeedforwardrdquo version of the problem rapid categorization
Movie courtesy of Jim DiCarlo
Biederman
1972 Potter 1975 Thorpe et al 1996
SHOW RSVP MOVIE
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
Modified from (Gross 1998)
A model of the ventral stream which is also an algorithm
[software available online]
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
A ldquofeedforwardrdquo version of the problem rapid categorization
Movie courtesy of Jim DiCarlo
Biederman
1972 Potter 1975 Thorpe et al 1996
SHOW RSVP MOVIE
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
Modified from (Gross 1998)
A model of the ventral stream which is also an algorithm
[software available online]
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
Modified from (Gross 1998)
A model of the ventral stream which is also an algorithm
[software available online]
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
hellipsolves the problem (if mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Object recognition for computer vision personal historical perspective
200019951990Viol
a amp Jo
nes 2
001
LeCun
et al
1998
Sch
neide
rman
amp Kan
ade 19
98 R
owley
Balu
jaamp K
anad
e 1998
Brune
lliamp P
oggio
1993
Turk
amp Pen
tland
1991
Sung amp
Pog
gio 19
94
Belong
ieamp M
alik 20
02 A
rgaw
alamp R
oth 20
02 U
llman
e
al 20
02
Face detection
Face identification
Car detection
Pedestrian detection
Multi-class multi-objects
Digit recognitionPer
ona an
d coll
eagu
es 19
96-n
owMoh
an P
apag
eorg
iouamp P
oggio
1999
Amit a
nd G
eman
1999
Osuna
amp Giro
si19
97
Best CVPRrsquo07 paper 10 yrs ago
Ferg
us et
al 20
03
Kanad
e 1974
hellip
Beimer
amp Pog
gio 19
95
Schne
iderm
anamp K
anad
e 2000
Torra
lbaet
al 20
04
hellip Many more excellent algorithms in the past few yearshellip
hellip
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Examples Learning Object Detection Finding Frontal Faces
bull
Training Databasebull
1000+ Real 3000+ VIRTUAL
bull
500000+ Non-Face Pattern
Sung amp Poggio 1995
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
~10 year old CBCL computer vision work pedestrian detection system in Mercedes test car now
becoming a product (MobilEye)
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Object recognition in cortex Historical perspective
198019701960
V1 catV1 monkeyExstrastriate
cortexIT-STS
Schille
r amp Le
e 199
1
Gross
et al
1969
1990
Hubel
amp Wies
el 19
62
Hubel
amp Wies
el 19
65
Hubel
amp Wies
el 19
59
Hubel
amp Wies
el 19
77
Kobata
keamp T
anak
a 199
4
Unger
leide
r amp M
Ishkin
1982
Per
rett R
olls e
t al 1
982
Zeki
1973
Schwar
tz et
al 19
83
Desim
one e
t al 1
984
Logo
thetis
et al
1995
hellip Much progress in the past 10 yrs
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Some personal history First step in developing a model
learning to recognize 3D objects in IT cortex
Poggio amp Edelman 1990
Examples of Visual Stimuli
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
An idea for a module for view-invariant identification
Architecture that accounts for invariances to 3D effects (gt1 view needed to learn)
Regularization Network (GRBF)with Gaussian kernels
View Angle
VIEW- INVARIANT
OBJECT- SPECIFIC
UNIT
Prediction neurons become
view-tuned through learning
Poggio amp Edelman 1990
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Learning to Recognize 3D Objects in IT Cortex
Logothetis
Pauls
amp Poggio 1995
Examples of Visual Stimuli
After human psychophysics (Buelthoff Edelman Tarr Sinha hellip) which supports models based on view-tuned units
hellip physiology
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Recording Sites in Anterior IT
LUNLAT
IOS
STS
AMTSLATSTS
AMTS
Ho=0
Logothetis Pauls
amp Poggio 1995
hellipneurons tuned to faces are intermingled
nearbyhellip
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Neurons tuned to object views as predicted by model
Logothetis
Pauls
amp Poggio 1995
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
A ldquoView-Tunedrdquo IT Cell
12 7224 8448 10860 12036 96
12 24 36 48 60 72 84 96 108 120 132 168o o o o o o o o o o o o
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120
Distractors
Target Views60
spi
kes
sec
800 msec
-108 -96 -84 -72 -60 -48 -36 -24 -12 0-168 -120 oo o o o o o o o oo o
Logothetis
Pauls
amp Poggio 1995
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
But also view-invariant object-specific neurons (5 of them over 1000 recordings)
Logothetis
Pauls
amp Poggio 1995
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Scale Invariant Responses of an IT Neuron
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Sik
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
Spi
kes
sec
0
76
0 2000 3000Time (msec)
1000
ik
0
76
40 deg(x 16)
475 deg(x 19)
55 deg(x 22)
625 deg(x 25)
25 deg(x 10)
10 deg(x 04)
175 deg(x 07)
325 deg(x 13)
View-tuned cells scale invariance (one training view only) motivates present model
Logothetis
Pauls
amp Poggio 1995
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
From ldquoHMAXrdquo
to the model nowhellip
Riesenhuber amp Poggio 1999 2000 Serre Kouh
Cadieu
Knoblich
Kreiman amp Poggio 2005 Serre Oliva Poggio 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Neural Circuits
Source Modified from Jody Culhamrsquos
web slides
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Neuron basics
spikes
INPUT= pulses or graded potentials
COMPUTATION = Analog
OUTPUT = Chemical
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Some numbers
bull
Human Brainndash
1011-1012
neurons (1 million flies )
ndash
1014- 1015
synapses
bull
Neuronndash
Fundamental space dimension
bull
fine dendrites 01 micro
diameter lipid bilayer
membrane 5 nm thick specific proteins pumps channels receptors enzymes
ndash
Fundamental time length 1 msec
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
The cerebral cortex
Human
MacaqueThickness
3 ndash 4 mm
1 ndash 2 mmTotal surface area
~1600 cm2 ~160 cm2
(both sides)
(~50cm diam)
(~15cm diam)
Neurons mmsup2
~10⁵ mm2 ~ 10⁵ mm2 Total cortical neurons
~2 x 1010
~2 x 109
Visual cortex
300 ndash
500 cm2
80+cm2Visual Neurons
~4 x 109 ~109 neurons
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Gross Brain Anatomy
A large percentage of the cortex devoted to vision
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
The Visual System
[Van Essen amp Anderson 1990]
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
V1 hierarchy of simple and complex cells
LGN-type cells
Simple cells
Complex cells
(Hubel amp Wiesel 1959)
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
V1 Orientation selectivity
Hubel amp Wiesel movie
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
V1 Retinotopy
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
(Thorpe and Fabre-Thorpe 2001)
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Reproduced from [Kobatake amp Tanaka 1994] Reproduced from [Rolls 2004]
Beyond V1 A gradual increase in RF size
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Reproduced from (Kobatake
amp Tanaka 1994)
Beyond V1 A gradual increase in the complexity of the preferred stimulus
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
AIT Face cells
Reproduced from (Desimone et al 1984)
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
AIT Immediate recognition
Hung Kreiman Poggio amp DiCarlo 2005
identification
categorization
See also Oram
amp Perrett
1992 Tovee
et al 1993 Celebrini
et al 1993 Ringach
et al 1997 Rolls et al 1999 Keysers
et al 2001
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Source Lennie
Maunsell Movshon
The ventral stream
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
(Thorpe and Fabre-Thorpe 2001)
We consider feedforward architecture only
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Our present model of the ventral stream feedforward accounting only for ldquoimmediate
recognitionrdquo
bull
It is in the family of ldquoHubel-Wieselrdquo
models (Hubel amp Wiesel 1959 Fukushima 1980 Oram
amp Perrett 1993 Wallis amp Rolls 1997 Riesenhuber amp Poggio 1999 Thorpe 2002 Ullman
et al 2002 Mel 1997 Wersing
and Koerner 2003 LeCun
et al 1998 Amit
amp Mascaro
2003 Deco amp Rolls 2006hellip)
bull
As a biological model of object recognition in the ventral stream it is perhaps the most quantitative and faithful to known biology (though many detailsfacts are unknown or still to be incorporated)
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Two key computations
Unit types Pooling Computation Operation
Simple Selectivity
template matching
Gaussian-
tuning and-like
Complex Invariance Soft-max or-like
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Max-like operation (or-like)
Complex units
Gaussian-like tuning operation (and-like)
Simple units
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Gaussian tuning
Gaussian tuning in IT around 3D views
Logothetis
Pauls amp Poggio 1995
Gaussian tuning in V1 for orientation
Hubel amp Wiesel 1958
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Max-like operation
Max-like behavior in V1
Lampl
Ferster Poggio amp Riesenhuber 2004 see also Finn Prieber amp Ferster 2007
Gawne
amp Martin 2002
Max-like behavior in V4
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
(Knoblich Koch Poggio in prep Kouh amp Poggio 2007 Knoblich Bouvrie Poggio 2007)
Biophys implementation
bull
Max and Gaussian-like tuning can be approximated with same canonical circuit using shunting inhibition
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Of the same form as model of MT (Rust et al Nature Neuroscience 2007
Can be implemented by shunting inhibition (Grossberg
1973 Reichardt
et al 1983 Carandini
and Heeger 1994) and spike threshold variability (Anderson et al 2000 Miller and Troyer 2002)
Adelson
and Bergen (see also Hassenstein
and Reichardt 1956)
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
bull
Generic overcomplete dictionary of reusable shape
components (from V1 to IT) provide unique representation ndash
Unsupervised learning (from ~10000 natural images) during a developmental-like stage
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning ~ Gaussian RBF
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
S2 units
bull
Features of moderate complexity (n~1000 types)
bull
Combination of V1-like complex units at different orientations
bull
Synaptic weights w learned from natural images
bull
5-10 subunits chosen at random from all possible afferents (~100-1000)
stronger facilitation
stronger suppression
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Nature Neuroscience -
10 1313 -
1321 (2007) Published online 16 September 2007 | doi101038nn1975
Neurons in monkey visual area V2 encode combinations of orientationsAkiyuki Anzai Xinmiao Peng amp David C Van Essen
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
C2 units
bull
Same selectivity as S2 units but increased tolerance to position and size of preferred stimulus
bull
Local pooling over S2 units with same selectivity but slightly different positions and scales
bull
A prediction to be tested S2 units in V2 and C2 units in V4
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
A loose hierarchy
bull
Bypass routes along with main routes bull
From V2 to TEO (bypassing V4) (Morel amp Bullier 1990 Baizer et al 1991 Distler et al 1991 Weller amp Steele 1992 Nakamura et al 1993 Buffalo et al 2005)
bull
From V4 to TE (bypassing TEO) (Desimone et al 1980 Saleem et al 1992)
bull
ldquoReplicationrdquo
of simpler selectivities from lower to higher areas
bull
Richer dictionary of features with various levels of selectivity and invariance
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
bull
V1bull
Simple and complex cells tuning
(Schiller et al 1976 Hubel amp Wiesel 1965 Devalois et al 1982)bull
MAX operation in subset of complex cells (Lampl et al 2004)
bull
V4bull
Tuning for two-bar stimuli
(Reynolds Chelazzi amp Desimone 1999)bull
MAX operation
(Gawne et al 2002)bull
Two-spot interaction
(Freiwald et al 2005)bull
Tuning for boundary conformation (Pasupathy amp Connor 2001 Cadieu
et al 2007)bull
Tuning for Cartesian and non-Cartesian gratings
(Gallant et al 1996)
bull
ITbull
Tuning and invariance properties
(Logothetis et al 1995)bull
Differential role of IT and PFC in categorization
(Freedman et al 2001 2002 2003)bull
Read out data (Hung Kreiman Poggio amp DiCarlo 2005)bull
Pseudo-average effect in IT
(Zoccolan Cox amp DiCarlo 2005 Zoccolan Kouh Poggio amp DiCarlo 2007)
bull
Humanbull
Rapid categorization (Serre Oliva Poggio 2007)bull
Face processing (fMRI + psychophysics)
(Riesenhuber et al 2004 Jiang et al 2006)
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Comparison w| neural data
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Comparison w| V4
Pasupathy amp Connor 2001
Tuning for curvature and
boundary conformations
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
V4 neuron tuned to boundary conformations
ρ
= 078
Most similar model C2 unit
Pasupathy amp Connor 1999
No parameter fitting
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
J Neurophysiol 98 1733-1750 2007 First published June 27 2007
A Model of V4 Shape Selectivity and InvarianceCharles Cadieu Minjoon Kouh Anitha Pasupathy Charles E Connor Maximilian
Riesenhuber and Tomaso Poggio
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
V4 neurons (with attention directed away
from receptive field)
(Reynolds
et al 1999)
C2 unitsReference (fixed)
Probe (varying)
= response(probe) ndashresponse(reference)
= re
sp (p
air)
ndashre
sp (r
efer
ence
) Prediction Response of the pair is predicted to fall between the responses elicited by the stimuli alone
(Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005)
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Agreement w| IT Readout data
Hung Kreiman Poggio DiCarlo 2005
Serre Kouh Cadieu Knoblich Kreiman amp Poggio 2005
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Remarks
bull
The stage that includes (V4-PIT)-AIT-PFC represents a learning network of the Gaussian RBF type that is known (from learning theory) to generalize well
bull
In the theory the stage between IT and lsquorsquoPFCrdquo
is a linear classifier ndash
like the one used in the read-
out experimentsbull
The inputs to IT are a large dictionary of selective and invariant features
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Rapid categorization
SHOW ANIMAL NON_ANIMAL
MOVIE
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Database collected by Oliva amp Torralba
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Rapid categorization task (with mask to test feedforward model)
Animal presentor not
30 ms ISI
20 ms
Image
Interval Image-Mask
Mask1f noise
80 ms
Thorpe et al 1996 Van Rullen
amp Koch 2003 Bacon-Mace et al 2005
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
hellipsolves the problem (when mask forces feedforward processing)hellip
human-
observers (n = 24) 80
Model 82
Serre Oliva amp Poggio 2007
bull
drsquo~ standardized error rate bull
the higher the drsquo the better the perf
Human 80
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Further comparisons
bull
Image-by-image correlationndash
Heads ρ=071
ndash
Close-body ρ=084 ndash
Medium-body ρ=071
ndash
Far-body ρ=060
bull
Model predicts level of performance on rotated images (90 deg and inversion)
Serre Oliva amp Poggio PNAS 2007
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Source Bileschi amp Wolf
The street scene project
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
The StreetScenes Database
Object car pedestrian bicycle building tree road sky
Labeled Examples 5799 1449 209 5067 4932 3400 2562
3547 Images all taken with the same camera of the same type of scene and hand labeled with the same objects using the same labeling rules
httpcbclmitedusoftware-datasetsstreetscenes
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Examples
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Examples
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Examples
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Examples
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
bull
HoG (Dalal amp Triggs 2005)
bull
Part-based system (Leibe et al 2004)
bull
Local patch correlation (Torralba et al 2004)
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Serre Wolf Bileschi
Riesenhuber amp Poggio PAMI 2007
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
1
Problem of visual recognition visual cortex2
Historical background
3
Neurons and areas in the visual system4
Data and feedforward
hierarchical models
5
What is next
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
bull
Generic dictionary of shape components (from V1 to IT) ndash
Unsupervised learning during a developmental-like stage learning dictionaries of ldquotemplatesrdquo
at different S levels
bull
Task-specific circuits (from IT to PFC)ndash
Supervised learning
Layers of cortical processing units
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Learning the invariance from temporal continuity
w| T Masquelier amp S Thorpe (CNRS France)
Foldiak
1991 Perrett et al 1984 Wallis amp Rolls 1997 Einhauser
et al 2002 Wiskott amp Sejnowski 2002 Spratling 2005
Simple cells learn correlation in space (at the same time)
Complex cells learn correlation in time
SHOW MOVIE
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
What is next
bull
A challenge for physiology disprove basic aspects of the architecture
bull
Extensions to color and stereobull
More sophisticated unsupervised developmental learning in V1 V4 PIT how
bull
Extension to time and videosbull
Extending the simulation to integrate-and-fire neurons (~ 1 billion) and realistic synapses towards the neural code
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
The problemTraining Videos Testing videos
each video~4s 50~100 frames
bend jack jump
pjump run walk
side wave1 wave2
Dataset from (Blank et al 2005)
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
See also (Casile
amp Giese 2005 Sigala
et al 2005)
Previous work recognizing biological motion
using a model of the dorsal stream
Adapted from (Giese amp Poggio 2003)
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Baseline Our system
KTH Human 813 916
UCSD Mice 756 790
Weiz Human 867 963
Average 812 896
chances 10~20HH Jhuang T Serre L Wolf and T Poggio ICCV 2007
Multi-class recognition accuracy
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
What is next beyond feedforward models limitations
Serre Oliva Poggio 2007
Zoccolan
Kouh
Poggio DiCarlo 2007
Reynolds Chelazzi amp Desimone 1999
PsychophysicsV4
IT
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
What is next beyond feedforward models
limitations
bull
Recognition in clutter is increasingly difficultbull
Need for attentional bottleneck (Wolfe 1994) perhaps in V4 (see Gallant and Desimone and models by Walther + Serre)
bull
Notice this is a ldquonovelrdquo
justification for the need of attention
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
model
20 ms SOA (ISI=0 ms)
80 ms SOA (ISI=60 ms)
50 ms SOA (ISI=30 ms)
no mask condition
(Serre Oliva and Poggio PNAS 2007)
Limitations beyond 50 ms model not good enough
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Ongoing workhellip
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Attention and cortical
feedbacks
Model implementation of Wolfersquos guided search (1994)
Parallel (feature-based top- down attention) and serial
(spatial attention) to suppress clutter (Tsostos et al)
Chikkerur
Serre Walther Koch amp Poggio in prep
num items
dete
ctio
n ac
urac
y
Face detection scanning vs attention
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Example Results
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
What is next image inference backprojections and attentional mechanisms
bullbull
Normal recognition by humans (for long times) is Normal recognition by humans (for long times) is much much betterbetter
bullbull
Normal vision is Normal vision is much much more than categorization more than categorization or identification image or identification image understandinginferenceparsingunderstandinginferenceparsing
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
bullbull
FeedforwardFeedforward
model + model + backprojectionsbackprojections
implementing implementing featuralfeatural
and the and the spatial attention may improve recognition performance spatial attention may improve recognition performance
bullbull
BackprojectionsBackprojections
also accessroute information infrom lower areas to also accessroute information infrom lower areas to specific taskspecific task--dependent routines in PFC () Open questions dependent routines in PFC () Open questions
----
Which biophysical mechanisms for routinggating Which biophysical mechanisms for routinggating
----
Nature of routines in higher areas (Nature of routines in higher areas (egeg
PFC)PFC)
Attention-based models with high-level specialized routines
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Analysis-by-synthesis models eg
probabilistic inference in the ventral stream neurons represent conditional probabilities of the bottom-up sensory inputs given the top-down hypothesis and converge to globally consistent values
Bayesian models
Lee and Mumford 2003 Dean2005 Rao 2004 Hawkins 2004 Ullman 2007 Hinton 2005
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
What is next
bull
Image inference attentional
or Bayesian models
bull
Why hierarchies Beyond a model towards a theory
bull
Against hierarchies and the ventral stream subcortical
pathways (Bar et al 2006 hellip)
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
How then do the learning machines described in the theory compare with brains
One of the most obvious differences is the ability of people and animals to learn from very few examples
A comparison with real brains offers another related challenge to learning theory The ldquolearning algorithmsrdquowe have described in this paper correspond to one-layer architectures Are hierarchical architectures with more layers justifiable in terms of learning theory
Why hierarchies For instance the lowest levels of the hierarchy may represent a dictionary of features that can be shared across multiple classification tasks
There may also be the more fundamental issue of sample complexity Thus our ability of learning from just a few examples and its limitations may be related to the hierarchical architecture of cortex
Notices of the American Mathematical Society (AMS) Vol 50 No 5537-544 2003
The Mathematics of Learning Dealing with DataTomaso
Poggio
and Steve Smale
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Formalizing the hierarchy towards a theory
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Smale S T Poggio A Caponnetto and J Bouvrie Derived Distance towards a mathematical theory of visual cortex
CBCL Paper Massachusetts Institute of Technology Cambridge MA November 2007
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
From a model to a theory math results on unsupervised learning of invariances and of a dictionary of shapes
from image sequences
hellip
Caponetto Smale
and Poggio in preparation
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
(Obvious) caution remark
There is still much to do before we understand visionhellip
and the brain
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber
Collaborators
Comparison w| humans
A Oliva
Action recognition
H Jhuang
Attention
S Chikkerur
C Koch
D Walther
Computer vision
bull S Bileschi
bull L Wolf
Learning invariances
bullT Masquelier
bullS Thorpe
T Serre
Model
A Oliva
C Cadieu
U Knoblich
M Kouh
G Kreiman
M Riesenhuber