Generative
Gaussian Discriminant Analysis DA
FO X
x
Learning Algorithms
Generative
Discriminative Algorithms Naive Bayes
X X
x
x
to
O
o
Discriminative
Learn
or learn
pcy In how
n y Oz
Generative Learning Learn pcnly
Algorithm
PG
D.pe
PCn
pCnly o
suppose NEIR drop no _I convention
Assume plenty is Gaussian
zNpiE an
EEZ
Cov2 IE
Bayes
A
feature
Rule
Class
ply 1 n
pCn pCuly D.pCg D
pcniy
pCy o Gaussian Discriminant Analysis GDA
2,22 Trh Rmn fE T
Zn
fr
Gyu
2 MY ECT
ECD T I
Effi
PG expftzCxATE x p pCn1y o 1 expf x MOJE x p
Engh Elk
q expf tzcx p.IE l
pcxty D
parameters Mo Mi Z
G MD
Rh Rn
to9 PlyD
Cx y Jim likelihood
112Wh
Eco I
ply
Training
set
Joint
LQpromiE ITIpx g 10MoMi
Z
pcy
In
conditional likelihood
Maximum likelihood Eselmation
Et Philly
Discriminative
L
Max
E
It
ply
o
O
Mo fly
l lol no ¦Ì e log Ll
IEy M
No Igg oy X I gileo
EmIy 13 m
4
I grassy x
n
covariance
x
true
c
L I false3 0
sum of feature vectors
for examples with y O
wth
T
examples
yo
¦Ì II
E.tt
3gZ
tm
Xox peg
Mgm
IEEE gu Xx
AT
G
A
pyo arg may
of
Mo
c0 oO0
Prediction
arg mag PCgin
eg miz 250
pCx1y pcg pcn
argmyE25 O
2
For
Logistic Regression Mo Mi 2
Comparison
to
fixed
plot ply L Ix
ol Mo Mi as
PXlyLpiE Pylol Txjdm.MS
E a finof X
n0OOXXXXXXf
XXXxJQEhEs
os
y.tk
as
xxxskgmoldfnl
Generative
GDA assumes nlyo Npro E
Discriminative
Logistic Regression
ply414 eogf.IE
11
my4 NCA y n Berk
On
4 stronger
assumption
1
weaker
assumption
PG 1 u
n
I
top
e
I
o
yn Naive
Poisson AD Poisson
BerCp
Bayes
nly Hy
is
logistic
Feature vectors aar
u a
agrdwolff
CS229 zymurgy
mi
b
10,000
I
Ily
X E 0 sign
I f word i appears in email
wantto model
Xi
pcnly
ply
N
210,000
Assume his are conditionally independent given y
219000
possible values of
I
P X1 Xlgoooly P X ly P Ml Xi y
Past mile g P Nioooo
Placooooly
map n ly CS228 Graphical models
PCazly Plusly
ft ly PCR11y
slyo Play Ily
D
IT Maggy
Parameters
D
o
dy
My
likelihood
Clogdoily 17hPCayjoyly
Joint
MLE
I
iymdjly
y 7IELyL
EE
fog
EE
gil
giving