EMFT ne.ro ANalysis
INTRODUCE generalEM Algorithm
Gaussian mixture as Em Factor Analysis
RECT
WE Examiners
GMMs light soures
X CRd sans Koo
Give
PLZ j for it n j i k
called Z A latent variable not oserves meetly
Doj
find
WE
Tobrtyi Deceive Gmm Algo tu more general keepi E STEP M STER
way
Guess zcis I
ft parameters ofMonel
LATENT
variable morsel
n p Data
Lfo
i dog Plex j m
parameters
flat out varcaney
Picturealournegorithi Ccf
allotty
fat i our Estimate
At t
w
i
logP X tZ
zjo plz zjo Gmm
a
Argmax HopE
O
lot Easier to optimize
Leto
Not
1 lo 2 Lol
Ello
locweebound LOH Eighty
LESTEO find M STEP At
f
than thot
given G whydoes this Abstract
Argmax 410 Gmm Next How do WE find LtCol given 0
IDEA_ WE go tErm m dog Plx ja 5yd term
PLx je log 2 Plex z o
let Q ez St Ez Q cz l Q cz 30 EH
log
log
Just detalE
eos
Ei
EfI
can swithe log ECT 3 login SIDEBAR nextpage
iE logPixiE
EQtlogP Z Z Q Cz
forAny
Call this ELBO X Q sa QLz7do Plx2
Symbolposhi
Z
Qg yz PLx
Recall JENSEN WE
ii
Q Above
So Pick Any Q
gives AN 2 10
QCE
this holds And for Each term
Stown Phx jo 7 ELBO x Q o te Properly I
Property
Gort’s Dick Q sit I
WE pick A specific Q depending on
log
Q a logplxciyzcis.jot Q’i t
Pex Ei SETQciz Pzx
acts
isoaypcxa.fIajswtonoeI
0kt
e
Los C
do Q’izz p
Log C
So
Xci 4 that LIFT Low
Ct
no
there
us
dog C
ERHSDECBO
Plx 3z Q Lz
Q Lz
Pix 2
pza.
PLZ
ii
y
X
log
i
RESMEEM
STEP forEI n SET Qilz PLZ X 044
I STEP At Co2 Argfnax
Argfax EI ELBOW Q o
JENSENRenwnr.ee logCE x
SIDEBAR
3 E logled log is
concave function
TO
help
f
I1I Zb
a wprob1 you remember inequality thoughtexperiment x b w Rob 1 X
f AD
logCx
Since 2 Is like Efx Xa th Hb
Picture says for any
f Elzy IE FLED for flog 2 cha tCcHb