Upload
brainmoleculemarketing
View
735
Download
6
Tags:
Embed Size (px)
DESCRIPTION
The vanishing central executive - Distributed neural mechanisms of decision-making
Citation preview
The vanishing central executive
Distributed neural mechanisms of decision-making
The vanishing central executive
Distributed neural mechanisms of decision-making
Paul CisekPaul Cisek
Summer School in Cognitive SciencesSummer School in Cognitive SciencesEvolution and Function of ConsciousnessEvolution and Function of Consciousness
July 4, 2012July 4, 2012
Our question:Our question:
• When, where, how and why — since the origin of life on Earth When, where, how and why — since the origin of life on Earth about 4 billion years ago — did organisms' input/output about 4 billion years ago — did organisms' input/output functions become conscious input/output functions?functions become conscious input/output functions?
• Why “input/output” functions?Why “input/output” functions?
But first, another question:But first, another question:
What is behavior?What is behavior?
“The whole neural organism, it will be remembered, is, physiologically considered, but a machine for converting stimuli into reactions” (James, 1890, p. 372).
BehaviorBehavior: An analysis of the world, : An analysis of the world, followed by deliberation and planning, followed by deliberation and planning, followed by execution of the plan.followed by execution of the plan.
““sense, think, act”sense, think, act”
motoroutput
sensoryinput
ActionPerception
representationof the world
Cognition
representationof the motor plan
William James
Psychological architecture for behaviorPsychological architecture for behavior
motoroutput
sensoryinput
ActionPerception
representationof the world
Cognition
representationof the motor plan
• University courses• Textbooks• Journals• Conferences• Academic departments• Grant review committees• Scientists• Questions we ask• Theories we propose
Q: From where does this view originate?
“Dualism”“Dualism”
• Philosophy: The mind is non-physicalPhilosophy: The mind is non-physical– This forces interfaces between non-
physical mind and physical world
• Psychology: Study of the psychePsychology: Study of the psyche– Structuralism: The mind is studied
through introspection
Perception Actionsensoryinput
motoroutputMind
Socrates
Descartes
John Locke
Wilhelm Wundt
BehaviorismBehaviorism
– Stop this metaphysical nonsense…
Perception Actionsensoryinput
motoroutputMind
John Watson
BehaviorismBehaviorism
– Stop this metaphysical nonsense… – Perception and Action are directly linked– Subject matter: Learning laws which establish the linkage
Perception Actionsensoryinput
motoroutput
John Watson Ivan Pavlov Edward Thorndike B.F. Skinner
CognitivismCognitivism
– Internal processes are indispensable
Perception Actionsensoryinput
motoroutput
Tolman
CognitivismCognitivism
– Internal processes are indispensable– Cognition takes the mind’s place– A fully physical process – but what
kind?– “Information processing”
• Definition of “information”• Definition of “processing”
– Cognition is a computational process• Linguistics• Language of thought
Perception Actionsensoryinput
motoroutput
Cognition
Tolman
Shannon
Turing
Chomsky
Fodor
Computational view of the brainComputational view of the brain
• The “computer metaphor”The “computer metaphor”– Cognition is like computation:
Rule-based manipulation of representations (Newell & Simon, Pylyshyn)
– The mind is the software (Block)
– Studies of mental phenomena may be conducted independently of studies of brain physiology
• Less to worry about• Not so much known (yet) about the brain• Historical separation between psychology and biology
Newell & Simon
What kinds of representations?What kinds of representations?
• ““Descriptive” representationsDescriptive” representations– Capture knowledge about the world
and the organism– Explicit– Objective, accurate to external
reality, uncontaminated by internal states
– Examples:• Reconstructed visual image• 3-D map of the world• Labeled objects• Desired path of the hand in space
Descriptive representations
delineate the conceptual borders
between the processes that
construct them and the
processes that use them.
David Marr
input/output
^
input/output
^
Can we use this to understand the brain?Can we use this to understand the brain?
• Cognitive NeuroscienceCognitive Neuroscience– How are psychological /
cognitive functions produced by the brain?
– Ex: Decision-making– Based on the concepts of
cognitivism• Computation, descriptive
representations, working memory, attentional filters, motor programs, etc.
Michael Gazzaniga
Where is the central representation?Where is the central representation?
• The visual systemThe visual system– Two visual processing streams:
• ventral “what” • dorsal “where”
– Separate regions analyze color, motion, form, etc.
– Separate regions for near and far space
• Binding problemBinding problem– How to create the unified
representation of the world that is needed as input for cognition?
where
what
• Primary sensory and motor regionsPrimary sensory and motor regions
• ““Association” regionsAssociation” regions– Appear to first encode sensory,
then motor representations– Even true at the level of
individual neurons
Example: Lateral intraparietal areaExample: Lateral intraparietal area– is it “attention”?
(input to cognition)– or “intention”?
(output of cognition)– How could it be both?– Could it be cognition?
Perception, Cognition, & Action Systems?Perception, Cognition, & Action Systems?
Cognition
Cognition
Cognition?
• Decision-makingDecision-making– Neural correlates in prefrontal
and orbitofrontal cortex– also in parietal cortex– Premotor cortex– Supplemental motor area– Frontal eye fields– Basal ganglia– Even the superior colliculus
– Activity reflects decision everywhere at about the same time (~150ms)
Where is the central executive?Where is the central executive?
Conceptual challengesConceptual challenges
• The binding problemThe binding problem– How to create the unified representation of the world that is needed
as input for cognition?
• The problem of meaningThe problem of meaning– How does a computational process know the meaning of the
representations that it manipulates?– “Chinese Room” (Searle)– The “symbol grounding problem” (Harnad)– Representations are purely syntactic, they have no intrinsic
semantics, no meaning to the system that uses theminput/output
^
• Some observations:Some observations:1. This model inherits its structure from
mind-body dualism2. Was designed to explain the abstract
problem-solving behavior of adult humans3. Its concepts were developed under the
explicit assumption that the substrate doesn’t matter
• Perhaps it should not be surprising that this model has difficulty Perhaps it should not be surprising that this model has difficulty explaining neural data…explaining neural data…
Psychological architecture for behaviorPsychological architecture for behavior
motoroutput
sensoryinput
ActionPerception Cognition
EvolutionEvolution
• Two key concepts:Two key concepts:
– Natural selection• What is the selective advantage of X?
– Descent with modification• What are the phylogenetic origins of X?
Darwin
The long history of behaviorThe long history of behavior
Are brains input/output devices?Are brains input/output devices?
• What else could they be?What else could they be?
What kinds of devices are living systems?What kinds of devices are living systems?
• Control systemsControl systems::– Ex: Biochemistry
• Suppose there is some substance A necessaryfor survival
• Suppose there’s a catalyst for creating A whose action is regulated inversely by the concentration of A
• Feedback control system• Exploits consistencies in the laws of chemistry• Control loop within the organism: “Physiology”
A
What kinds of devices are living systems?What kinds of devices are living systems?
• Control systems can extend beyond the skinControl systems can extend beyond the skin– Ex: Kinesis
• Suppose substance B cannot be produced within the body, must be absorbed from the world
• If the local concentration of substance B falls below desired levels, move randomly
• Exploits statistics of nutrient distributions(assumes that there is more elsewhere)
• Control loop that extends outside the skin: “Behavior”
– Reliable motor-sensory contingencies exist• Statistics of food distributions (move → find food)• Laws of optics and mechanics (contract muscle → arm moves)• Laws of interaction (you show teeth → I back off)
• Animals are constantly doing whatever brings them to the most Animals are constantly doing whatever brings them to the most desirable situation (full stomach, safety, etc.)desirable situation (full stomach, safety, etc.)
• “Behavior: The control of perception” (Powers, 1973)
AB
Concentration of [B]
Different ways of looking at behaviorDifferent ways of looking at behavior
1.1. Given a perception, produce the best actionGiven a perception, produce the best action
“The whole neural organism, it will be remembered, is, physiologically considered, but a machine for converting stimuli into reactions” (James, 1890).
2.2. Of the possible actions, produce that which Of the possible actions, produce that which results in the best perceptionresults in the best perception
“What we have is a circuit… the motor response determines the stimulus, just as truly as sensory stimulus determines movement” (Dewey, 1896).
John Dewey
William James
EthologyEthology
• Studies of animal behavior in Studies of animal behavior in the wildthe wild
• Species-specific behavioral Species-specific behavioral nichesniches
• ““Closed-loop” sensorimotor Closed-loop” sensorimotor controlcontrol
• Key stimuliKey stimuli
Von Uexküll
Lorenz & Von Holst
Tinbergen
What kinds of representations?What kinds of representations?
• ““Descriptive” representationsDescriptive” representations– Capture knowledge about the world
and the organism– Explicit– Objective, accurate to external
reality, uncontaminated by internal states
– Examples:• Reconstructed visual image• 3-D map of the world• Labeled objects• Desired path of the hand in space
• ““Pragmatic” representationsPragmatic” representations– Used to guide interaction between
the world and the organism– Implicit– Subjective, mix external reality and
internal state, often correlate with many variables at once
– Examples:• Salience map• Motor signals to the limb• Subject-dependent opportunities for
action (“affordances”)
David Marr J.J. Gibson
Example: Decision-makingExample: Decision-making
What to do?What to do?Move the queen?Protect the pawn?Threaten the
knight?
““Selection”Selection”
How to do it?How to do it?Which grasp point?What trajectory?How to avoid
obstacles?
““Specification”Specification”
• Classical model:Classical model:– First decide what to do (select) then plan the movement (specify)– Sense, think, act
Decision-making in the wildDecision-making in the wild
• The world presents animals with multiple opportunities for action (“affordances”)The world presents animals with multiple opportunities for action (“affordances”)• Cannot perform all actions at the same timeCannot perform all actions at the same time• Real-time activity is constantly modifying affordances, introducing new ones, etc.Real-time activity is constantly modifying affordances, introducing new ones, etc.
Action specification and selection must occur in parallel
Sensorimotor contingencies influence how selection should be done
Cel
l act
ivity
Direction
Dis
tanc
e
Cel
l act
ivity
Specification and selection in parallelSpecification and selection in parallel
• Action SpecificationAction Specification: Activation of parameter regions corresponding : Activation of parameter regions corresponding to potential actionsto potential actions
• Action SelectionAction Selection: Competition between distinct regions of activity: Competition between distinct regions of activity
A population of tuned neurons
What are the neural substrates?What are the neural substrates?
attention Specification in the dorsal visual stream– Cells sensitive to spatial visual information
(Ungerleider & Mishkin …)– Involved in action guidance (Milner &
Goodale)– Divergence into separate sub-streams, each
specialized toward different kinds of actions (Stein; Andersen; Colby & Goldberg; Matelli & Luppino ...)
– An increasing influence of attentional effects, enhancing information from particular regions of interest (Duncan & Desimone; Posner & Gilbert; Treue; Boynton ...)
– Parietal representation of external world is “sparse” (Goldberg)
Specification in the dorsal visual stream– Cells sensitive to spatial visual information
(Ungerleider & Mishkin …)– Involved in action guidance (Milner &
Goodale)– Divergence into separate sub-streams, each
specialized toward different kinds of actions (Stein; Andersen; Colby & Goldberg; Matelli & Luppino ...)
– An increasing influence of attentional effects, enhancing information from particular regions of interest (Duncan & Desimone; Posner & Gilbert; Treue; Boynton ...)
– Parietal representation of external world is “sparse” (Goldberg)
potential actions
attention
Fronto-parietal systemFronto-parietal system– Activity related to potential motor
actions (Andersen; Georgopoulos; Kalaska; Wise; Hoshi & Tanji)
– Competition between potential actions– Various biasing factors
• attention (Goldberg; Steinmetz)• behavioral relevance
(Mountcastle; Seal & Gross)• probability (Glimcher; Shadlen)• reward (Glimcher; Olson)
attention
behavioralbiasing
potential actions
Basal gangliaBasal ganglia– Cortico-striatal-pallido-thalamo-cortical
loops (Alexander; Middleton & Strick)– Selection of actions from among
alternatives (Mink; Redgrave et al.)– Reward (Hikosaka; Schultz)
cognitivedecision-making
attention
behavioralbiasing
objectidentity
potential actions
Prefrontal cortexPrefrontal cortex– High-level decisions
based on knowledge about object identity (Fuster; Miller; Tanji…)
– Receives ventral stream information on object identity (Sakata…)
cognitivedecision-making
objectidentity
attention
behavioralbiasing
potential actions
ExecutionExecution– Fast visual feedback
(Prablanc; Desmurget)– Forward models
(Ito; Wolpert; Miall)
motorcommand
visual feedback
predictedfeedback
cognitivedecision-making
objectidentity
attention
behavioralbiasing
potential actions
motorcommand
visual feedback
predictedfeedback
cognitivedecision-making
objectidentity
attention
behavioralbiasing
potential actions
motorcommand
visual feedback
predictedfeedback
“Affordance competition hypothesis”“Affordance competition hypothesis”
• Continuous specification of currently available potential actionsContinuous specification of currently available potential actions
• Competition between potential action representations in fronto-Competition between potential action representations in fronto-parietal regionsparietal regions
• Biasing from frontal and subcortical areasBiasing from frontal and subcortical areas
• Decision is made through a “distributed consensus”Decision is made through a “distributed consensus”
Cisek (2007) Cisek (2007) Phil.Trans.Royal Soc. B.Phil.Trans.Royal Soc. B.
Behavior
Behavior
Actionspecification
Actionselection
Perception Cognition Action
auditionvision
planning
attention forwardmodels
inversekinematics
decisionmaking
trajectorygeneration
visionof space
objectrecognition
propositionallogic
reinforcementlearning
actionsequencing
proprioception
attention
forwardarm models
inverse armkinematics
decisionmaking
visionof nearby
space
objectrecognition
reaching
key stimulusdetection action
sequencing
propositionallogic
grasping running affect
reinforcementlearningproprio-
ception
PredictionsPredictions
• Multiple potential actions can be specified simultaneouslyMultiple potential actions can be specified simultaneously
• Biased competition between potential actionsBiased competition between potential actions
• Everything occurs in parallelEverything occurs in parallel
Neural activity specifies multiple actionsNeural activity specifies multiple actions
Classic model:Classic model:– Store information, decide,
then plan one action
Affordance competition:Affordance competition:– Specify both actions,
then select one
Cell PD
Cell PD
TimeTime
RostralRostralPMdPMd
CaudalCaudalPMdPMd
PrimaryPrimaryMotorMotorCortexCortex
Cisek & Kalaska (2005) Neuron
PredictionsPredictions
• Multiple potential actions can be specified simultaneouslyMultiple potential actions can be specified simultaneously
• Biased competition between potential actionsBiased competition between potential actions
• Everything occurs in parallelEverything occurs in parallel
Biased choice taskBiased choice task
1-TARGET
CHT
2-TARGET
CHT
GO
FREE
67%
FORCED
GO
33%
DELAY
GO
THT
Reward: 1
THT
Reward: 1
DELAY
2 drops
3 drops
1 drop
THT
Reward: 3
AlexandrePastor-Bernier
Neural activity in premotor cortexNeural activity in premotor cortex
• No effect of valueNo effect of valuein 1T taskin 1T task
• However, if another However, if another target is present, then target is present, then activity activity increasesincreases with with value of preferred value of preferred targettarget
• If value of preferred If value of preferred target is constant, target is constant, activity activity decreasesdecreases with with value of other targetvalue of other target
• Activity decreases with Activity decreases with distance between distance between targetstargets
Pastor-Bernier & Cisek (2011) J. Neurosci.
Distance-dependent interactionsDistance-dependent interactions
• More activity when targets are More activity when targets are closercloser
• Compare the strength of the Compare the strength of the competition as a function of competition as a function of target distancetarget distance– As distance increases, slope is
increasingly negative
• The competition is strongest The competition is strongest between cells with the largest between cells with the largest difference in preferred directionsdifference in preferred directions
Why should it matter that distance matters?Why should it matter that distance matters?
• The distance effect suggests that the decision is madeThe distance effect suggests that the decision is madewithin the sensorimotor systemwithin the sensorimotor system– If decisions were purely cognitive (“I prefer to get 3 drops of juice
over 1 drop”), then they should be determined in an abstract space– The dynamics of the competition which determines choice depend
on the spatial relationship between the movements themselves
PredictionsPredictions
• Multiple potential actions can be specified simultaneouslyMultiple potential actions can be specified simultaneously
• Biased competition between potential actionsBiased competition between potential actions
• Everything occurs in parallelEverything occurs in parallel
cognitivedecision-making
objectidentity
attention
behavioralbiasing
potential actions
motorcommand
visual feedback
predictedfeedback
cognitivedecision-making
objectidentity
attention
behavioralbiasing
potential actions
motorcommand
visual feedback
predictedfeedback
TimingTiming
• An animal is constantly interacting with the worldAn animal is constantly interacting with the world– Continuous sensorimotor control of ongoing actions– Continuous specification of alternative actions– Continuous evaluation of value– Continuous tradeoffs between persisting in a given activity or
switching to a different, currently available one
• Specification and selection must normally occur in parallelSpecification and selection must normally occur in parallel
• However, if we put the animal in the labHowever, if we put the animal in the lab– Time is broken into discrete “trials” each of which begins with a
stimulus and ends with a response– The stimulus is deliberately made independent from the response
• What should we see?What should we see?
cognitivedecision-making
objectidentity
attention
behavioralbiasing
potential actions
motorcommand
visual feedback
predictedfeedback
potential actions
Wave 1. Fast feedforward sweepWave 1. Fast feedforward sweep
cognitivedecision-making
objectidentity
attention
behavioralbiasing
potential actions
Wave 2. Attentional/Decisional modulationWave 2. Attentional/Decisional modulation
Two waves of activityTwo waves of activity
• Measured LFPs from various regions of cerebral cortexMeasured LFPs from various regions of cerebral cortex
• Monkeys performed a conditional GO / NOGO taskMonkeys performed a conditional GO / NOGO task
Ledberg et al. (2007) Cerebral Cortex
Two waves of activityTwo waves of activity
1.1. Fast feedforward sweepFast feedforward sweep• Activation in ~50ms
throughout dorsal stream and frontal cortex
2.2. Attentional/DecisionalAttentional/Decisional• About 150ms post-
stimulus, discrimination of Go/Nogo throughout the cortex
Summary 1: Experimental dataSummary 1: Experimental data
• Simultaneous specification of multiple potential actionsSimultaneous specification of multiple potential actions– Arm reaching system (PMd, PRR, M1)– Grasping system (AIP, PMv)– Saccade system (LIP, FEF, Superior colliculus)
• Biased competitionBiased competition– Potential actions compete against each other within sensorimotor
maps, influenced by a variety of biasing factors (e.g. reward)– NOTE: Similar mechanism as attention (Duncan & Desimone)
• “Attention” is selection near sensors, “decision” is selection near effectors
– Influences depend on geometry – decisions are not simply abstract• These are “pragmatic” representations, not “descriptive”
– Decision is made through a “distributed consensus”
• Parallel specification and selection systemsParallel specification and selection systems
Summary 2: Theoretical conceptsSummary 2: Theoretical concepts
• ““Affordance competition hypothesis”Affordance competition hypothesis”– Instead of serial Perception, Cognition, & Action modules, we have parallel
specification and selection systems– Better match to neural data– Better suited to the kinds of tasks that dominated animal behavior
• ““Pragmatic representations”Pragmatic representations”– Neural activity aimed not at describing the world, but at mediating
interaction with the world– Correlation with external and internal variables is necessary, but mixtures
are useful (e.g. spatial direction mixed with reward values)– Conjecture: Most, but not all, neural activity is of this kind
• “Descriptive” representations (e.g. in the ventral stream) emerged in evolution as specializations of pragmatic representations for advanced selection
• Cognitive advances evolved through hierarchical elaborationCognitive advances evolved through hierarchical elaboration– Diversification of fronto-parietal loops, cortico-striatal circuits, cortico-
cerebellar circuits, into progressively anterior/abstract systems– Interaction lays the foundation for cognition (Piaget)
Summary 3: Philosophical implicationsSummary 3: Philosophical implications
• There is no central executiveThere is no central executive– Decisions emerge through a distributed consensus
• Classic problems in a different contextClassic problems in a different context– Binding problem:
• Activity of separate streams is coherent by virtue of dealing with the same world
– Symbol grounding problem:• Interaction has meaning by virtue of influencing the variables critical for life• Symbols are specializations (“shorthand notation”) that emerged late in evolution,
already within the context of grounded interaction
– The “Hard” problem• Feeling is different than doing
– Being inside the loop is different than observing it from the outside– Private language, beetle in box, squirrel in head, 1st person perspective, the “Umwelt”
• The computer metaphorThe computer metaphor– With all due respect to Alan Turing, the computer metaphor is misleading as
a model for the brain– What matters is control (Wiener, Ashby, Powers, Gibson, Dewey, etc.)
““The great end of life is not The great end of life is not knowledge but action”knowledge but action”
– T. H. Huxley (1825-1895)
““Your head is there to move Your head is there to move you around”you around”
– R.E.M. (1980-2011)
THANK YOUTHANK YOU
• Lab membersLab members– Marie-Claude Labonté– Alexandre Pastor-Bernier– David Thura– Ignasi Cos– Matthew Carland– Jessica Trung
• AlumniAlumni– Jean-Philippe Thivierge– Thomas Michelet– Valeriya Gritsenko
FOUNDATIONTHEEJLBFOUNDATIONTHEEJLB