Mutt by T Apply separable fat III x E cut I'coin ation at y ut y X G - - PDF document

SLIDE 1

Summary fromprevious lecture

Data Za

Zn

Pf

j

B

• e 0

loss Estimation

z

t

A regularize

minimize

Into In flzi o 111101

Example

Sparse regression Lasso Zi Gi Xi Xi N lo Id

Yi OIXitti

Lnlol

InHy X

0112 21014

• f

Algorithms

Ily riot I

Gradientdescent

qt hot seVLr.LI

step size Prox gradient Acc gradient Mirror descent FOM

commonstructure Zi

Gi xi

Yi ER

ER

llzijd lcyi.TK

en

t _In Illyi

Exit theCy

Xo

Ten lol

XTfG Xo

tGiXo

tG io

fcyn.x.io

fGi5I 2gecyy

q

gt se XTfly XOt

Mutt by T

in

Apply separable fat

SLIDE 2

III

x E cut

at y

I'coin ation

ut

X G con Ot f

G'Icu

ut y

at c R

OtcRd

Ft

R

R

Can

we analyze GFOMs

Find the optimal

• ne

statist

Setting

Xi

• Nlo kn

8 Eco N

Yi

HATE

wi

wi yaw

h suffrey

QF

Nooo Oe

PLIED

iii

atone

IT

Tim

For any

GFOM

linguist

11910.1123

where

explicit

Further there exists

a special GFOM

BayesAMD

hn.hofsa.IT l95amp 0olP II Proof

n

Reduction GFOM 7 AMP

2

Sharp analysis

n d a

3

BAMP

• ptimal among AMP

1

Phaseretrieval

Z

Oo

Yi

Cti

Ei

Ln

101 7 yi sxi.o.JP

Xz

Tpo

SLIDE 3

l

Tx

Lula

Lala

Illy Xolighth

Ot

z

gt

X

z

t a

It

y Ige

a

actin 2 Eit

t 12

AMI

M

Ot

zlottxiit at

ut

y

Xot zItn

7

• The

Soo.iq LawL 0

tie't

Z NcoD

ind of 0

O

fELIN0 142 del 0324 N

PO

I

SE

• nto.IE

II.CEti ooiT sEEo tttZ OD E

11010.11

ELCHAEZIA OT to

4107,81

E4COtttZ.O l

2 otf EL zlo.tt2in

07

ex

g

r

f RIENZI

d

It

z

Nco 41

GAUSSIAN

Can improve over soft thr AMP

get

ht Ott X ut

ut

y

Xut

Stat

II

07

ELKE0

2

B

yes AMP

SLIDE 4

ht yl

EL 0 I

itEE y

T'tt

O't f

mm se

II

µ

tee Z

mm set't

ELL0

E 10 22 172

TIZ

J

g mmse.at

L

10

1

T

• e

Kf

f

110.18 golf

poet E 8otes

e

we

EASY

8

401

• f

initthool E

y

Is

8

T

E

HARD

t

fixed

nd

linear c.pk

10

f

yi

ai 055

L

spectral init

SLIDE 5

O

l

rel

i

l

Old

iterations

Oclogn

go

cordhdom init

H FI Kyi

xi

v

M 80

c

v

M