Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
TopicsPerceptronExponential FamilyGeneralized Linear Models
Softmax RegressionMulticlass Classification
Logistic Regressiona
It
g 2 z GG L 230O Z CO
hope 1home g E'aIt Ota
0 i O T x y how ngdifferent
y how0 algo got it rightL if wrong y L EEEy wrong y 0
EnzoQ
OEl
iO O
O
want 0 70 but 0Th LOO OtLn tLu5 n On th Nu
Exponential FamiliesPDF
plugin blog exp Thy acne
y data beg exp 2TTG
n natural parameterTcg sufficient statistic
y in classblog Base measure
acq log partition function
y scalar
q vector scalar
TG vector scalar I match
b g scalar
ExampleBernoulli Binary Dataprobability of event
p g lo 494t y
exp log 994 09 exp y tog lo ta g logo a
exp log 9 t logo 0
Tks Fat
bly ITcg gn log to q Csigmoid
acq logCto log l teenlog Iten
Gaussian w fixed variance or L
ply a En exp HIIe 5k expCpu y Zai
Tk Tk Tanbcg fan expC ETG yq µacne of IzProperties natural paramsMLE w r t 2 is concave
negative laglikelihood NLL is convex
EEg n a IqacnVarly y zigzag
GLMAssumptions Design Clonolces
c yl K O Exponential familyReal GaussianBinary Bernoullicount PoissonRT Gamma ExponentialDist peta Dirichlet BayesianMystats
Il if 0Th O EIRDk E Rd
cu Test tune output ECyIn o
how IE yl x o
hoax log ply OTH
Framxi Ip
Learning Update RulehoCa
Rest
Oji o t x y hoax ajplug in appropriate holu
Terminologyq natural parameter
µ EE y R g y canonical response for
2 g A canonical link f n
god znach3 paramatrizations
Modelparameters Natural param CanonicalParamBernoulli0 Y µ ot Gaussian
pi Entg X Poisson
Designgtlearning choice
Logistic Regressionholn Elgin o 9 n teen
Assumption'Legression
ammo A Fundo
Classification
xx xxxx x x0.5FEEI.eefzxxxxaxc xx
a
n1He 2
sigmoidLota of
o
Softmax RegressionMulticlass Classification
Cross Entropy minimization
U2 g 00OO
00Lie
K E IRDy E 0,13 one hot vector 0 0,40
d
Glass C IRD f oOz
class c o o D 3 kg in i e oooo
IO l
IEnno
OOI ht.ee
l ll l
Goal a
oin
normalize
PG
t 4
goal i mm distance between 2 dumba
man Cross Entropy pp I ply tog pacyYC 0,0 03log fly
00Thlog e
E eOEn
NgCEEQa o y
Fund Oo Oo 8D Gradient Descent