Probabilistic Graph Models WMCI 20154

90
Probabilistic Reasoning using Graph Models Vineet Sahula Dept. of ECE, MNIT Jaipur [email protected]

description

.

Transcript of Probabilistic Graph Models WMCI 20154

Page 1: Probabilistic Graph Models WMCI 20154

Prob

abili

stic

Rea

soni

ng u

sing

Gra

ph

Mod

els

Vine

etSa

hula

Dept.

of E

CE, M

NIT

Jaipu

rsa

hula@

acm

.org

Page 2: Probabilistic Graph Models WMCI 20154

Outli

ne

•Pr

obab

ilistic

reas

oning

unde

r unc

ertai

nty•

Infer

ence

s in g

raph

mod

els–

Baye

sian,

Marko

v Ran

dom

Fields

•Re

ason

ing w

ith tim

e•

Case

stud

y–

Nano

scale

memo

ry mo

delin

g: Re

liabil

ity &

Life

time

evalu

ation

in pr

esen

ce of

erro

rs•

Conc

lusion

s

WM

SC 1

6-Ja

n-20

15

Page 3: Probabilistic Graph Models WMCI 20154

ACKN

OWLE

DG

MEN

TS

Slide

s in p

arts

borro

wed f

rom

-Pr

of. W

elling

, UCI

rvine

, CA

-Pr

of. R

. J. M

oone

y, Ut

exas

, Aus

tin WM

SC 1

6-Ja

n-20

15

Page 4: Probabilistic Graph Models WMCI 20154

Perfo

rman

ce E

valu

atio

n of

SoC

Com

mun

icat

ion

Arch

itect

ures

-

Ulha

sD

eshm

ukh

•De

velop

ment

of eff

icien

t per

forma

nce e

valua

tion &

desig

n spa

ce

explo

ratio

n fra

mewo

rk for

comm

unica

tion

arch

itectu

re a

t Sys

tem

level

(ESL

)•

Cons

idere

d stoc

hasti

c natu

re of

Mod

el pa

rame

ters-

GENE

RALIZ

ED di

stribu

tions

for P

E’s c

ompu

tation

time,

and

comm

. Tim

e etc.

•Mo

del s

uppo

rts co

ncur

renc

y as w

ell as

hier

arch

y-Bu

s, mu

ltiple-

hiera

rchica

l bus

, NoC

•An

alytic

al me

thod

for ev

aluati

on o

f Gen

erali

zed

Semi

Mar

kov

Proc

ess (

GSMP

)-AF

OME,

HCF

G, S

AN•

BW, W

aiting

time,

utiliz

ation

facto

r, qu

eue l

ength

comp

uted

WM

SC 1

6-Ja

n-20

15

Page 5: Probabilistic Graph Models WMCI 20154

Effic

ient

Fac

ial f

eatu

re re

cogn

ition

-R. A

. Pat

il•

Prop

osed

an al

gorith

m for

facia

l featu

re ex

tracti

on w

ith tit

led fa

ces,

& no

n-ne

utra

l exp

ress

ion o

n ini

tial fr

ame

•Au

tomate

d wire

mod

el fitt

ing•

Topo

grap

hical

maps

(neu

ral c

ompu

tation

) for

map

ping o

f featu

re

vecto

rs, an

d mult

iclas

s SVM

•Ha

rdwa

re m

appin

g for

effic

ient, r

eal ti

me op

erati

on-s

ystol

ic ar

ray

comp

utatio

n, pa

rtial re

confi

gura

tion f

or lo

w po

wer

WM

SC 1

6-Ja

n-20

15

Page 6: Probabilistic Graph Models WMCI 20154

Nan

osca

le M

emor

y M

odel

ing

& S

ynth

esis

-R. K

umaw

at•

Molec

ular D

evice

Mod

el co

nside

red f

or su

b-sy

stem

mode

l gen

erati

on•

Nano

cellb

ased

Mole

cular

Mem

ory S

ynthe

sis in

pres

ence

of h

ard

erro

rs–

Omnip

otent

Train

ing–

Morta

l Tra

ining

•Pr

obab

ilistic

Ana

lysis

of Na

noce

llMole

cular

Mem

ory

–Re

liabil

ity P

redic

tion

of Na

noce

llin

•Sp

atial

Doma

in•

Time D

omain

–Ex

tende

d Con

tinuo

us P

aram

eter B

irth-D

eath

Mode

l for P

roba

bilist

ic An

alysis

of N

anoc

ellin

pres

ence

of T

rans

ient E

rrors.

WM

SC 1

6-Ja

n-20

15

Page 7: Probabilistic Graph Models WMCI 20154

Gra

phic

al M

odel

s

A ‘

mar

riag

e’ b

etw

een

prob

abili

ty th

eory

and

gra

ph th

eory

Why

pro

babi

litie

s?

•R

easo

ning

with

unc

erta

intie

s, co

nfid

ence

leve

ls•

Man

y pr

oces

ses a

re in

here

ntly

‘noi

sy’

robu

stne

ss is

sues

Why

gra

phs?

•Pr

ovid

e ne

cess

ary

stru

ctur

e in

larg

e m

odel

s:

-Des

igni

ng n

ew p

roba

bilis

tic m

odel

s.-R

eadi

ng o

ut (c

ondi

tiona

l) in

depe

nden

cies

.

•In

fere

nce

& o

ptim

izat

ion:

-Dyn

amic

al p

rogr

amm

ing

-Bel

ief P

ropa

gatio

n

Page 8: Probabilistic Graph Models WMCI 20154

Type

s of

Gra

phic

al M

odel

Und

irec

ted

grap

h (M

arko

v ra

ndom

fie

ld)

Dir

ecte

d gr

aph

(Bay

esia

n ne

twor

k)∏

∏=

iij

ji

iji

ix

xx

Zx

P)

()

()

,(

)(

1)

ψ

i

j

)(

ii

),

( )(

ji

ijx

)|

()

()

(pa

rent

s∏

=i

ii

xx

Px

P

i

Pare

nts(

i)

fact

or g

raph

s

inte

ract

ions

vari

able

s

Page 9: Probabilistic Graph Models WMCI 20154

9

Gra

phic

al M

odel

s

•If n

o ass

umpti

on of

inde

pend

ence

is m

ade,

then a

n exp

onen

tial

numb

er of

para

meter

s mus

t be e

stima

ted fo

r sou

nd pr

obab

ilistic

inf

eren

ce.

•No

reali

stic a

moun

t of tr

aining

data

is su

fficien

t to e

stima

te so

man

y pa

rame

ters.

•If a

blan

ket a

ssum

ption

of co

nditio

nal in

depe

nden

ce is

mad

e, eff

icien

t train

ing an

d infe

renc

e is p

ossib

le, bu

t suc

h a st

rong

as

sump

tion

is ra

rely

warra

nted.

•Gr

aphi

cal m

odels

use d

irecte

d or u

ndire

cted g

raph

s ove

r a se

t of

rand

om va

riable

s to e

xplic

itly sp

ecify

varia

ble d

epen

denc

ies a

nd

allow

for le

ss re

strict

ive in

depe

nden

ce as

sump

tions

whil

e lim

iting

the nu

mber

of pa

rame

ters t

hat m

ust b

e esti

mated

.–

Baye

sian

Netw

orks

: Dire

cted a

cycli

c gra

phs t

hat in

dicate

caus

al str

uctur

e.–

Mark

ov N

etwo

rks:

Undir

ected

grap

hs th

at ca

pture

gene

ral d

epen

denc

ies.

Page 10: Probabilistic Graph Models WMCI 20154

10

Baye

sian

Net

wor

ks

•Di

recte

d Acy

clic G

raph

(DAG

)–

Node

s are

rand

om va

riable

s–

Edge

s ind

icate

caus

al inf

luenc

es

Bur

glar

yE

arth

quak

e

Ala

rm

John

Cal

lsM

aryC

alls

Page 11: Probabilistic Graph Models WMCI 20154

11

Cond

ition

al P

roba

bilit

y Ta

bles

•Ea

ch no

de ha

s a co

nditi

onal

prob

abilit

y tab

le(C

PT) t

hat g

ives t

he

prob

abilit

y of e

ach

of its

value

s give

n ev

ery p

ossib

le co

mbina

tion

of va

lues f

or its

pare

nts (c

ondit

ioning

case

).–

Roots

(sou

rces)

of the

DAG

that

have

no pa

rents

are g

iven p

rior p

roba

bilitie

s.

Bur

glar

yE

arth

quak

e

Ala

rm

John

Cal

lsM

aryC

alls

P(B)

.001

P(E)

.002

BE

P(A

)T

T.9

5T

F.9

4F

T.2

9F

F.0

01

AP(

M)

T.7

0F

.01

AP(

J)T

.90

F.0

5

Page 12: Probabilistic Graph Models WMCI 20154

12

CPT

Com

men

ts

•Pr

obab

ility o

f false

not g

iven s

ince r

ows m

ust a

dd to

1.•

Exam

ple re

quire

s 10 p

aram

eters

rathe

r tha

n 25 –

1 = 31

for s

pecif

ying t

he fu

ll join

t dis

tributi

on.

•Nu

mber

of pa

rame

ters i

n the

CPT

for a

node

is ex

pone

ntial

in the

numb

er of

pare

nts

(fan-

in).

Page 13: Probabilistic Graph Models WMCI 20154

13

Join

t Dis

tribu

tions

for B

ayes

Net

s

•A

Baye

sian N

etwor

k imp

licitly

defin

es a

joint

distrib

ution

.))

(Pa

rent

s|

()

,...

,(

12

1i

n ii

nX

xP

xx

xP

∏ =

=

•E

xam

ple

)(

EB

AM

JP

¬∧

¬∧

∧∧

)(

)(

)|

()

|(

)|

(E

PB

PE

BA

PA

MP

AJ

¬¬

∧¬

=00

062

.099

8.0

999

.000

1.0

7.09.0

××

×=

•T

here

fore

an

inef

fici

ent a

ppro

ach

to in

fere

nce

is:

–1)

Com

pute

the

join

t dis

trib

utio

n us

ing

this

equ

atio

n.–

2) C

ompu

te a

ny d

esir

ed c

ondi

tiona

l pro

babi

lity

usin

g th

e jo

int d

istr

ibut

ion.

Page 14: Probabilistic Graph Models WMCI 20154

14

Naï

ve B

ayes

as

a Ba

yes

Net

•Na

ïve B

ayes

is a

simple

Bay

es N

et Y

X1

X2

…X

n

•Pr

iors

P(Y

) an

d co

nditi

onal

s P(

Xi|Y

) fo

r N

aïve

Bay

es p

rovi

de C

PTs

for

the

netw

ork.

Page 15: Probabilistic Graph Models WMCI 20154

Con

stru

ctin

g B

ayes

ian

Net

wor

ks

Cho

ose

the

right

ord

er fr

om c

ause

s to

effe

cts.

P(x1

,x2,

…,x

n) =

P(x

n|xn

-1,..

,x1)

P(xn

-1,…

,x1)

= Π

P(x

i|xi-1

,…,x

1)

--ch

ain

rule

Exam

ple:

P(x1

,x2,

x3) =

P(x

1|x2

,x3)

P(x2

|x3)

P(x3

)

Page 16: Probabilistic Graph Models WMCI 20154

How

to c

onst

ruct

BN

P(x1

,x2,

x3)

x3

x2

x1

root

cau

se

leaf

Cor

rect

ord

er: a

dd ro

ot c

ause

s firs

t, an

d th

en

“lea

ves”

, with

no

influ

ence

on

othe

r nod

es.

Page 17: Probabilistic Graph Models WMCI 20154

Com

pact

ness

BN

are

loca

lly st

ruct

ured

syst

ems.

They

repr

esen

t joi

nt d

istri

butio

ns c

ompa

ctly

.

Ass

ume

nra

ndom

var

iabl

es, e

ach

influ

ence

dby

k n

odes

. Si

ze B

BN

: n2k

F

ull s

ize:

2n

Page 18: Probabilistic Graph Models WMCI 20154

18

Inde

pend

enci

es in

Bay

es N

ets

•If r

emov

ing a

subs

et of

node

s Sfro

m the

netw

ork r

ende

rs no

des X

iand

Xjd

iscon

necte

d, the

n Xia

nd X

jare

ind

epen

dent

given

S, i.

e. P(

X i| X j, S

) = P

(Xi|

S)•

Howe

ver,

this i

s too

stric

t a cr

iteria

for c

ondit

ional

indep

ende

nce s

ince t

wo no

des w

ill sti

ll be c

onsid

ered

ind

epen

dent

if the

ir sim

ply ex

ists s

ome v

ariab

le tha

t de

pend

s on b

oth.

–Fo

r exa

mple,

Bur

glary

and E

arthq

uake

shou

ld be

cons

idere

d ind

epen

dent

since

they

both

caus

e Alar

m.

Page 19: Probabilistic Graph Models WMCI 20154

19

Inde

pend

enci

es in

Bay

es N

ets

(con

t.)

•Un

less w

e kno

w so

methi

ng ab

out a

comm

on ef

fect o

f two “

indep

ende

nt ca

uses

” or a

desc

ende

nt of

a com

mon e

ffect,

then

they

can b

e co

nside

red

indep

ende

nt.

–Fo

r exa

mple,

if we

know

nothi

ng el

se, E

arthq

uake

and B

urgla

ry ar

e ind

epen

dent.

•Ho

weve

r, if w

e hav

e info

rmati

on a

bout

a com

mon e

ffect

(or d

esce

nden

t the

reof)

then

the t

wo “in

depe

nden

t” ca

uses

bec

ome

prob

abilis

ticall

y lin

ked s

ince e

viden

ce fo

r one

caus

e can

“exp

lain a

way”

the ot

her.

–Fo

r exa

mple,

if we

know

the a

larm

went

off th

at so

meon

e call

ed ab

out th

e alar

m,

then i

t mak

es ea

rthqu

ake a

nd bu

rglar

y dep

ende

nt sin

ce ev

idenc

e for

earth

quak

e de

creas

es be

lief in

burg

lary.

and v

ice ve

rsa.

Page 20: Probabilistic Graph Models WMCI 20154

Why

do

we

need

it?

•A

nsw

er q

ueri

es :

-Giv

en p

ast p

urch

ases

, in

wha

t gen

re b

ooks

is a

clie

nt in

tere

sted

?-G

iven

a n

oisy

imag

e, w

hat w

as th

e or

igin

al im

age?

•L

earn

ing

prob

abili

stic

mod

els

from

exa

mpl

es

( exp

ecta

tion

max

imiz

atio

n, it

erat

ive

scal

ing

)

•Opt

imiz

atio

n pr

oble

ms:

min

-cut

, max

-flo

w, V

iterb

i, …

Infe

renc

e in

Gra

phic

al M

odel

s

Infe

renc

e:•

Ans

wer

que

ries

abo

ut u

nobs

erve

d ra

ndom

var

iabl

es, g

iven

val

ues

of o

bser

ved

rand

om v

aria

bles

.

•M

ore

gene

ral:

com

pute

thei

r jo

int p

oste

rior

dist

ribu

tion: (

|)

{(

|)}

iPu

oor

Pu

o

lear

ning

infe

renc

e

Page 21: Probabilistic Graph Models WMCI 20154

Appr

oxim

ate

Infe

renc

e

Infe

renc

e is

com

puta

tiona

lly in

trac

tabl

e fo

r la

rge

grap

hs (

with

cyc

les)

.

App

roxi

mat

e m

etho

ds:

•M

arko

v C

hain

Mon

te C

arlo

sam

plin

g.

•M

ean

fiel

d an

d m

ore

stru

ctur

ed v

aria

tiona

ltec

hniq

ues.

•B

elie

f Pr

opag

atio

n al

gori

thm

s.

Page 22: Probabilistic Graph Models WMCI 20154

Belie

f Pro

paga

tion

on tr

ees

ik

k

k

k

ij

k

k

kMki

∏∑

→→

∝k

ii

kx

ii

ji

ijj

ji

xM

xx

xx

Mi

)(

)(

),

()

ψ

Com

patib

ilitie

s (i

nter

actio

ns)

exte

rnal

evi

denc

e

∏∝

kk

ki

ii

ix

Mx

xb

)(

)(

)(

ψ

mes

sage

belie

f (

appr

oxim

ate

mar

gina

l pro

babi

lity)

Page 23: Probabilistic Graph Models WMCI 20154

Belie

f Pro

paga

tion

on lo

opy

grap

hs

ik

k

k

k

ij

k

k

kMki

∏∑

→→

∝k

ii

kx

ii

ji

ijj

ji

xM

xx

xx

Mi

)(

)(

),

()

ψ

Com

patib

ilitie

s (i

nter

actio

ns)

exte

rnal

evi

denc

e

∏∝

kk

ki

ii

ix

Mx

xb

)(

)(

)(

ψ

mes

sage

belie

f (

appr

oxim

ate

mar

gina

l pro

babi

lity)

Page 24: Probabilistic Graph Models WMCI 20154

Som

e fa

cts

abou

t BP

•B

P is

exa

ct o

n tr

ees.

•If

BP

conv

erge

s it

has

reac

hed

a lo

cal m

inim

um o

f an

obj

ectiv

e fu

nctio

n (t

he B

ethe

fre

e en

ergy

Yed

idia

et.a

l ‘00

, H

eske

s’0

2)of

ten

good

app

roxi

mat

ion

•If

it co

nver

ges,

con

verg

ence

is f

ast n

ear

the

fixe

d po

int.

•M

any

exci

ting

appl

icat

ions

: -

erro

r co

rrec

ting

deco

ding

(Mac

Kay

, Yed

idia

, McE

liece

, Fre

y)-

visi

on (

Fre

eman

, Wei

ss)

-bi

oinf

orm

atic

s (W

eiss

)-

cons

trai

nt s

atis

fact

ion

prob

lem

s (D

echt

er)

-ga

me

theo

ry (

Kea

rns)

-…

Page 25: Probabilistic Graph Models WMCI 20154

Gen

eral

ized

Bel

ief P

ropa

gatio

n

∏∑

→∝ k

ii

kx

ii

ji

ij

jj

i

xM

xx

x

xM

i

)(

)(

),

(

)(

ψψ

Idea

: To

gues

s th

e di

stri

butio

n of

one

of

your

nei

ghbo

rs, y

ou a

sk

your

oth

erne

ighb

ors

to g

uess

you

r di

stri

butio

n. O

pini

ons

get

com

bine

d m

ultip

licat

ivel

y.

∏∑

→∝ k

ii

kx

ii

ji

ij

jj

i

xM

xx

x

xM

i

)(

)(

),

(

)(

ψψ

BP

GB

P

Page 26: Probabilistic Graph Models WMCI 20154

26

Baye

s N

et In

fere

nce

•Gi

ven k

nown

value

s for

some

evid

ence

varia

bles

, dete

rmine

the

poste

rior p

roba

bility

of so

me q

uery

varia

bles

.•

Exam

ple: G

iven t

hat J

ohn c

alls,

what

is the

prob

abilit

y tha

t the

re is

a Bu

rglar

y?

Bur

glar

yE

arth

quak

e

Ala

rm

John

Cal

lsM

aryC

alls

???

John

cal

ls 9

0% o

f the

tim

e th

ere

is a

n A

larm

and

the A

larm

det

ects

94%

of B

urgl

arie

s so

peop

lege

nera

lly th

ink

it sh

ould

be

fairl

y hi

gh.

How

ever

, thi

s ign

ores

the

prio

rpr

obab

ility

of J

ohn

calli

ng.

Page 27: Probabilistic Graph Models WMCI 20154

27

Baye

s N

et In

fere

nce

•Ex

ample

: Give

n tha

t Joh

n call

s, wh

at is

the pr

obab

ility t

hat

there

is a

Burg

lary?

Bur

glar

yE

arth

quak

e

Ala

rm

John

Cal

lsM

aryC

alls

???

John

als

o ca

lls 5

% o

f the

tim

e w

hen

ther

eis

no

Ala

rm. S

o ov

er 1

,000

day

s we

expe

ct 1

Bur

glar

y an

d Jo

hn w

ill p

roba

bly

call.

How

ever

, he

will

als

o ca

ll w

ith a

fa

lse

repo

rt 50

tim

es o

n av

erag

e. S

o th

e ca

ll is

abo

ut 5

0 tim

es m

ore

likel

y a

fals

e re

port:

P(B

urgl

ary

| Joh

nCal

ls) ≈

0.0

2

P(B)

.001

AP(

J)T

.90

F.0

5

Page 28: Probabilistic Graph Models WMCI 20154

28

Baye

s N

et In

fere

nce

•Ex

ample

: Give

n tha

t Joh

n call

s, wh

at is

the pr

obab

ility t

hat

there

is a

Burg

lary?

Bur

glar

yE

arth

quak

e

Ala

rm

John

Cal

lsM

aryC

alls

???

Act

ual p

roba

bilit

y of

Bur

glar

y is

0.0

16

sinc

e th

e al

arm

is n

ot p

erfe

ct (a

n Ea

rthqu

ake

coul

d ha

ve se

t it o

ff or

it

coul

d ha

ve g

one

off o

n its

ow

n). O

n th

e ot

her s

ide,

eve

n if

ther

e w

as n

ot a

n al

arm

and

John

cal

led

inco

rrec

tly, t

here

co

uld

have

bee

n an

und

etec

ted

Bur

glar

y an

yway

, but

this

is u

nlik

ely.

P(B)

.001

AP(

J)T

.90

F.0

5

Page 29: Probabilistic Graph Models WMCI 20154

29

Type

s of

Infe

renc

e

Page 30: Probabilistic Graph Models WMCI 20154

30

Sam

ple

Infe

renc

es

•Di

agno

stic

(evid

entia

l, abd

uctiv

e): F

rom

effec

t to ca

use.

–P(

Burg

lary |

John

Calls

) = 0.

016

–P(

Burg

lary |

John

Calls

∧Ma

ryCall

s) =

0.29

–P(

Alar

m | J

ohnC

alls ∧

MaryC

alls)

= 0.7

6–

P(Ea

rthqu

ake |

John

Calls

∧Ma

ryCall

s) =

0.18

•Ca

usal

(pre

dict

ive):

From

caus

e to e

ffect

–P(

John

Calls

| Bur

glary)

= 0.

86–

P(Ma

ryCall

s | B

urgla

ry) =

0.67

•In

terc

ausa

l (ex

plain

ing

away

): Be

twee

n ca

uses

of a

comm

on e

ffect.

–P(

Burg

lary |

Alar

m) =

0.37

6–

P(Bu

rglar

y | A

larm ∧

Earth

quak

e) =

0.00

3•

Mixe

d: T

wo or

mor

e of th

e abo

ve co

mbine

d –

(diag

nosti

c and

caus

al) P

(Alar

m | J

ohnC

alls ∧

¬Ear

thqua

ke) =

0.03

–(d

iagno

stic a

nd in

terca

usal)

P(B

urgla

ry | J

ohnC

alls ∧

¬Ear

thqua

ke) =

0.01

7

Page 31: Probabilistic Graph Models WMCI 20154

31

Com

plex

ity o

f Bay

es N

et In

fere

nce

•In

gene

ral, t

he pr

oblem

of B

ayes

Net

infer

ence

is N

P-ha

rd

(exp

onen

tial in

the s

ize of

the g

raph

).•

For s

ingl

y-co

nnec

ted

netw

orks

or p

olyt

rees

in wh

ich

there

are n

o und

irecte

d loo

ps, th

ere a

re lin

ear-t

ime

algor

ithms

base

d on b

elief

pro

paga

tion.

–Ea

ch n

ode

send

s loc

al ev

idenc

e me

ssag

es to

their

child

ren

and

pare

nts.

–Ea

ch n

ode u

pdate

s beli

ef in

each

of its

poss

ible v

alues

base

d on

inco

ming

mes

sage

s fro

m it n

eighb

ors a

nd pr

opag

ates

evide

nce o

n to i

ts ne

ighbo

rs.•

Ther

e are

appr

oxim

ation

s to i

nfere

nce f

or ge

nera

l ne

twor

ks ba

sed o

n loo

py b

elief

pro

paga

tion

that

itera

tively

refin

es pr

obab

ilities

that

conv

erge

to ac

cura

te va

lues i

n the

limit.

Page 32: Probabilistic Graph Models WMCI 20154

32

Belie

f Pro

paga

tion

Exam

ple

•λm

essa

ges a

re se

nt fro

m ch

ildre

n to p

aren

ts re

pres

entin

g abd

uctiv

e evid

ence

for a

node

.•

πme

ssag

es ar

e sen

t from

pare

nts to

child

ren

repr

esen

ting c

ausa

l evid

ence

for a

node

.

Bur

glar

yE

arth

quak

e

Ala

rm

John

Cal

lsM

aryC

alls

λ

λλ

πA

larm

Bur

glar

yE

arth

quak

e

Mar

yCal

ls

Page 33: Probabilistic Graph Models WMCI 20154

33

Mul

tiply

Con

nect

ed N

etw

orks

•Ne

twor

ks w

ith un

direc

ted lo

ops,

more

than

one d

irecte

d pa

th be

twee

n som

e pair

of no

des.

•In

gen

eral

, inf

eren

ce in

suc

h ne

twor

ks is

NP-

hard

.

•So

me

met

hods

con

stru

ct a

pol

ytre

e(s)

fro

m g

iven

ne

twor

k an

d pe

rfor

m in

fere

nce

on tr

ansf

orm

ed g

raph

.

Page 34: Probabilistic Graph Models WMCI 20154

34

Nod

e Cl

uste

ring

•El

imina

te all

loop

s by m

ergin

g nod

es to

crea

te m

egan

odes

that

have

the c

ross

-pro

duct

of va

lues o

f the

mer

ged n

odes

.

•N

umbe

r of

val

ues

for

mer

ged

node

is e

xpon

entia

l in

the

num

ber

of n

odes

mer

ged.

•St

ill r

easo

nabl

y tr

acta

ble

for

man

y ne

twor

k to

polo

gies

req

uiri

ng r

elat

ivel

y lit

tle m

ergi

ng to

el

imin

ate

loop

s.

Page 35: Probabilistic Graph Models WMCI 20154

35

Stat

istic

al R

evol

utio

n

•Ac

ross

AI th

ere h

as be

en a

move

ment

from

logic-

base

d ap

proa

ches

to ap

proa

ches

base

d on p

roba

bility

and s

tatist

ics.

–St

atisti

cal n

atura

l lang

uage

pro

cess

ing–

Stati

stica

l com

puter

visio

n–

Stati

stica

l robo

t nav

igatio

n–

Stati

stica

l lear

ning

•Mo

st ap

proa

ches

are f

eatur

e-ba

sed a

nd “p

ropo

sition

al” an

d do

not h

andle

comp

lex re

lation

al de

scrip

tions

with

mult

iple e

ntitie

s lik

e tho

se ty

picall

y req

uiring

pred

icate

logic.

Page 36: Probabilistic Graph Models WMCI 20154

Marko

v Ran

dom

Field

(MRF

)Ba

yesia

n Netw

ork (

BN)

Prob

abilit

y Tra

nsfer

Matr

ix (P

TM)

Prob

abilit

y Dec

ision

Diag

ram

(PDD

)Pr

obab

ilistic

Mod

el Ev

aluati

ons:

a com

para

tive s

tudy

Prop

osed

exte

nded

birt

h-de

ath

mod

el

PROB

ABIL

ISTI

C M

ODEL

ING

AP

PROA

CHES

FOR

NAN

OSCA

LE

SUB-

SYST

EMS W

MSC

16-

Jan-

2015

Page 37: Probabilistic Graph Models WMCI 20154

Expe

rimen

tal S

etup

for O

mni

pote

nt T

rain

ing

of M

olec

ular

Mem

ory

WM

SC 1

6-Ja

n-20

15

Page 38: Probabilistic Graph Models WMCI 20154

Algo

rithm

for O

mni

pote

nt T

rain

ing

ALGO

RITH

M 1 :

Omn

ipoten

t train

ing of

n-bit

mem

ory u

sing G

eneti

c Algo

rithm

Inpu

t: untra

ined_

MEM

: an u

ntrain

ed sa

mple

of a n

anoc

ell;

Initia

l_P: in

itial p

opula

tion

of ‘M

’ mole

cules

with

each

eithe

r ‘ON

’ or ‘O

FF’;

fitnes

s_M

EM():

fitne

ss fu

nctio

n for

n-bit

mem

ory;

/*It d

efine

s the

desir

ed ou

tput v

oltag

e lev

els fo

r Write

/Rea

d mo

des o

f n-b

it mem

ory *

/nu

m_G

EN: n

umbe

r of g

ener

ation

s;m

ax_G

EN: m

axim

um n

umbe

r of g

ener

ation

s;Ou

tput

:tra

ined_

MEM

: A tr

ained

n-b

it mem

ory d

evice

;St

eps:

1: Ini

tializ

e Gen

etic A

lgorith

m (G

A) w

ith po

pulat

ion in

itial_P

;2:

for n

um_G

EN=1

to m

ax_G

ENdo

(i): M

odify

the s

tate

of ea

ch m

olecu

le to

‘ON’

or ‘O

FF’ in

unt

raine

d_M

EM;

(ii): S

imula

te the

mod

ified u

ntra

ined

_ M

EM u

sing H

SPIC

E;(iii

) Stor

e ou

tput v

oltag

e (O

ut) v

alues

for R

ead a

nd W

rite op

erati

ons f

or al

l n bi

ts in

volta

ge_V

AL;

(iii):

The G

A ev

aluate

s fitne

ss fu

nctio

n fitn

ess_

MEM

() us

ing vo

ltage

_VAL

;(iv

): It

quan

tifies

the

fitnes

s of e

ach i

ndivi

dual

and g

ener

ates n

ew po

pulat

ion;

end

3: Th

e GA

conv

erge

s and

prov

ides a

s outp

ut:(i)

:the

optim

al se

t of

molec

ules,

each

with

state

‘ON’

or ‘O

FF’;

(ii):th

e tra

ined_

MEM

by us

ing th

ese c

ontro

l volt

age v

alues

;re

turn

traine

d_M

EM;

WM

SC 1

6-Ja

n-20

15

Page 39: Probabilistic Graph Models WMCI 20154

Sim

ulat

ion

Resu

lt of

100

0 sa

mpl

es o

f 1-b

it m

emor

y

WM

SC 1

6-Ja

n-20

15

Page 40: Probabilistic Graph Models WMCI 20154

Prop

osed

Arc

hite

ctur

e of

Nan

ocel

l bas

ed

Mul

ti-bi

t Mem

ory

WM

SC 1

6-Ja

n-20

15

Page 41: Probabilistic Graph Models WMCI 20154

Expe

rimen

tal S

etup

for M

orta

l Tra

inin

g of

M

olec

ular

Mem

ory

WM

SC 1

6-Ja

n-20

15

Page 42: Probabilistic Graph Models WMCI 20154

Algo

rithm

-Mor

tal T

rain

ing

ALGO

RITH

M 2 :

Mor

tal tr

aining

of n-

bit m

emor

y usin

g Gen

etic A

lgorith

mIn

put: un

traine

d_M

EM: a

n untr

ained

samp

le of

a nan

ocell

;In

itial_C

: initia

l pop

ulatio

n of r

ando

mly g

ener

ated

kCon

trol V

oltag

e Sign

als (C

VS);

/* Th

ese C

VS ar

e den

oted a

s Ci ,∀i= 1

to k

/* Ea

ch C

ivalu

e is e

ither

low

(Vlow

=0.5

V) o

r high

(Vhig

h=2.

0V)*/

fitnes

s_M

EM():

fitne

ss fu

nctio

n for

n-bit

mem

ory;

/*It d

efine

s the

desir

ed ou

tput v

oltag

e lev

els fo

r Write

/Rea

d mo

des o

f n-b

it mem

ory *

/nu

m_G

EN: n

umbe

r of g

ener

ation

s;m

ax_G

EN: m

axim

um n

umbe

r of g

ener

ation

s;Ou

tput

:tra

ined_

MEM

: A tr

ained

n-b

it mem

ory d

evice

;St

eps:

1: Ini

tializ

e Gen

etic A

lgorith

m (G

A) w

ith po

pulat

ion in

itial_C

;2:

for n

um_G

EN=1

to m

ax_G

ENdo

(i): M

odify

the v

alues

of C

ontro

l Volt

age s

ignals

in un

traine

d_M

EM;

(ii): S

imula

te the

mod

ified u

ntra

ined

_ M

EM u

sing H

SPIC

E;(iii

) Stor

e ou

tput v

oltag

e (O

ut) v

alues

for R

ead a

nd W

rite op

erati

ons f

or al

l n bi

ts in

volta

ge_V

AL;

(iii):

The G

A ev

aluate

s fitne

ss fu

nctio

n fitn

ess_

MEM

() us

ing vo

ltage

_VAL

;(iv

): It

quan

tifies

the

fitnes

s of e

ach i

ndivi

dual

and g

ener

ates n

ew po

pulat

ion;

end

3: Th

e GA

conv

erge

s and

prov

ides a

s outp

ut:(i)

:the

optim

al se

t of c

ontro

l volt

age

(Ci)

value

s;(ii)

:the

traine

d_M

EMby

using

thes

e con

trol v

oltag

e valu

es;

retur

n tra

ined_

MEM

;W

MSC

16-

Jan-

2015

Page 43: Probabilistic Graph Models WMCI 20154

Des

ign

Spac

e Ex

plor

atio

n fo

r Mor

tal T

rain

ing

ALGO

RITH

M 3 :

Des

ign S

pace

Exp

lorati

on fo

r Mor

tal tr

aining

of n

-bit m

olecu

lar m

emor

yIn

put: un

traine

d_M

EM: a

n untr

ained

samp

le of

a nan

ocell

cons

isting

of N

nano

partic

les;

num

_C: n

umbe

r of C

ontro

l Volt

age S

ignals

;nu

m_C

_ LO

C : n

umbe

r of n

anop

artic

les to

whic

h Con

trol V

oltag

e Sign

als ca

n be a

pplie

d;/*n

um_C

_ LO

C ⊂N */

/*num

_C_L

OC≤

(N−5

) */

/*num

_C_L

OC≥

num

C*/

Outp

ut:

traine

d_M

EM: s

et of

all su

cces

sfully

train

ed n

-bit m

emor

y dev

ices;

num

_MEM

: num

ber o

f train

ed n-

bit m

emor

y sam

ples i

n the

set tr

ained

_MEM

Step

s:

1: Ge

nera

te_C

(num

_C);

/*Gen

erate

a se

t Swh

ich co

nsist

s of a

ll bipa

rtition

s of n

um_C

Contr

ol Vo

ltage

Sign

als, C

i, ∀i=1 to n

um_C

.*/

/*Ass

ign {V

low=0

.5V

or V

high=

2.0

V} to

each

Ci j,∀i= 1

to n

um_C

, j =

1 to

size

of (

S). *

/ /*S

ave

these

in C

VS_V

ALUE

S.*/

;2:

Gene

rate

_C_L

oc(n

umC

LOC,

num

C);

/*Gen

erate

all p

ossib

le co

mbina

tions

of n

um_C

_LOC

nano

partic

les to

whic

h num

_Cco

ntrol

volta

ge si

gnals

can b

e con

necte

d. */

/*Sav

e the

se in

LOCA

TION

S.*/

;3:

Gene

rate

_Mem

(unt

raine

d_M

EM, L

OCAT

IONS

);/*M

odify

unt

raine

d_ M

EMby

using

Con

trol V

oltag

e Sign

al loc

ation

spec

ified i

n LOC

ATIO

NS */

/*Sav

e all

mem

ory s

ample

s in M

EM_M

ODIF

IED

*/ ;

4: Ge

nera

te_S

pfiles

(MEM

_MOD

IFIE

D, C

VS_V

ALUE

S);

/*For

eac

h mem

ory s

ample

in M

EM_M

ODIF

IED,

ass

ign C

ontro

l Volt

age

value

s fro

m CV

S_VA

LUES

.*//*S

ave

all ne

wly m

odifie

dmem

ory s

ample

s in S

P_FI

LES.

*/ ;

5: Si

mul

ate_

Mem

(SP_

FILE

S);

/*Sim

ulate

all m

emor

y sam

ples i

n SP_

FIL

ESus

ing H

SPIC

E. */

/*Sav

e ou

tput in

OUT

_FIL

ES.*/

;6:

Fetc

h_Me

m2b

it (O

UT_F

ILES

);/*S

elect

all th

ose

memo

ry sa

mples

, for w

hich n

-bit m

emor

y rea

d and

write

is do

ne in

acce

ptable

noise

mar

gin, u

sing t

he

OUT_

FILE

S.*/

/*Cop

y all s

electe

d mem

ory s

ample

s to

traine

d_M

EM. *

/;re

turn

traine

d_M

EM, n

um_M

EM;

WM

SC 1

6-Ja

n-20

15

Page 44: Probabilistic Graph Models WMCI 20154

Num

ber o

fna

nopa

rticle

s per

Na

noce

ll (N)

Num

ber o

f Mo

lecul

es(M

)

No. o

fsuc

cess

fully

train

ed n

anoc

ell in

stan

ces w

ith iC

VS

6 CVS

8C

VS

1

0 CVS

1

2 CVS

2097

20

00

2010

41

00

0

3022

327

135

0

3021

433

194

0

4039

910

355

132

4039

298

4310

1

Numb

er of

Mem

ory S

ample

s (c x

p)31

x 92

412

7 x 49

551

1 x 66

2047

x 1

Tabl

e 1.

Num

ber o

f suc

cess

fully

trai

ned

nano

cell

conf

igur

atio

n fo

r 2-b

it m

emor

y re

ad a

nd w

rite

oper

atio

n.

Her

e no

ise

mar

gin

for D

ata

read

ope

ratio

n on

out

put

volta

ge n

ode

is c

onsi

dere

d as

0.4

V.

WM

SC 1

6-Ja

n-20

15

Page 45: Probabilistic Graph Models WMCI 20154

Prob

abili

stic

Ana

lysi

s of

Nan

ocel

l in

Spat

ial

Dom

ain

Forc

orre

ctfu

nctio

ning

atle

asto

neof

the

min

imal

path

mus

tbe

pres

ent

betw

een

Inpu

tand

Out

putP

orts

.is

ane

cess

ary

cond

ition

.Th

epr

esen

ceof

mul

tiple

path

sw

illin

trodu

cere

dund

ancy

and

incr

ease

prob

abili

tyof

getti

ngco

rrec

tout

put.

Let

usm

odel

the

nano

cell

asa

plan

argr

aph

G(V

,E),

whe

rena

nopa

rticl

esar

eth

eno

des

and

mol

ecul

arsw

itche

sin

‘ON

’sta

tear

eth

eed

ges.

The

grap

hG

(V,E

)is

assu

med

tobe

adi

rect

iona

lgra

phsu

chth

atal

lthe

mol

ecul

esar

eor

ient

edin

sam

edi

rect

ion,

i.e.f

rom

inpu

tto

the

outp

ut.

WM

SC 1

6-Ja

n-20

15

Page 46: Probabilistic Graph Models WMCI 20154

•Th

epr

imar

yin

putp

orti

sth

ero

otno

dean

dth

epr

imar

you

tput

port

isth

ele

afve

rtex

ofth

egr

aph

G(V

,E).

•C

onsi

dera

nano

cell

with

‘N’n

anop

artic

les

and

‘M’m

olec

ular

switc

hes.

•A

ssum

eth

atth

ese

nano

parti

cles

are

alw

ays

pres

ent

and

mol

ecul

arsw

itche

sar

edi

strib

uted

byG

auss

ian

dist

ribut

ion

with

inth

ena

noce

ll.•

Am

olec

ular

switc

hii

spr

esen

tin

‘ON

’sta

tew

ithpr

obab

ility

p i.

•Th

ispr

obab

ility

that

ithmo

lecule

ispr

esen

tin

ONsta

teis

calle

dre

liabil

ityof

that

molec

uleat

thatin

stant

oftim

e,de

noted

asr(p

i)

WM

SC 1

6-Ja

n-20

15

Page 47: Probabilistic Graph Models WMCI 20154

•So

,the

prob

abilit

ytha

tthe

reis

noed

gein

betw

een

two

nano

partic

lesis

(1−

pi).A

gain,

thepr

obab

ilityt

hata

tleas

tone

edge

ispr

esen

tbetw

een

twon

odes

isgiv

enby

:

•Le

tuss

uppo

setha

tsma

llnan

ocell

cons

istso

fthr

eemo

lecula

rswi

tches

(m1,

m2an

dm3

)and

three

nano

partic

les(A

,Ban

dC)

,as

show

nin

Figur

ea.

•He

re,in

putv

oltag

eis

appli

edon

Aan

drec

eived

onC.

•Th

isca

nbed

one

viatw

omini

malp

aths:-

WM

SC 1

6-Ja

n-20

15

Page 48: Probabilistic Graph Models WMCI 20154

•Co

rrect

outpu

twill

bere

ceive

don

Cvia

path

ifboth

molec

ulesm

1an

dm2

are

pres

ent.

Simi

larly,

data

willb

eco

rrectl

yre

ceive

don

Cvia

path

ACif

themo

lecule

m3is

inON

state.

Thes

ear

ethe

two

redu

ndan

tpath

sand

atlea

ston

eof

them

shou

ldbe

worki

ngco

rrectl

yto

obtai

nco

rrect

outpu

t.Th

us,th

e•

Prob

abilit

yofr

eceiv

ingco

rrect

data

onno

deC

isgiv

enby

•P Pa

th=

(Pro

babil

itytha

tboth

m1an

dm2

are

pres

enti

n‘O

N’sta

teon

path

ABC)

or(P

roba

bility

thatm

3isp

rese

ntin

‘ON’

state

onpa

thAC

)

WM

SC 1

6-Ja

n-20

15

Page 49: Probabilistic Graph Models WMCI 20154

Eval

uatin

g re

liabi

lity

boun

ds fo

r exa

mpl

e ci

rcui

t

Miss

ing

Molec

ules

V(lo

w)V(

High

)Pa

th P

roba

bilit

y ( A

toC)

Uppe

r Bou

nd

Low

erBo

und

None

0.493

51.9

935

0.625

m10.4

951

1.995

10.5

00m2

0.495

11.9

951

0.500

m30.4

902

1.990

20.2

50m1

, m2

0.495

11.9

951

0.500

m2, m

30

00.0

00m1

, m3

00

0.000

m1, m

2, m3

00

0.000

Table

3. S

imula

tion R

esult

s for

an ex

ample

nano

cell c

onfig

urati

on w

hen

none

or so

me of

the m

olecu

les ar

e miss

ing

WM

SC 1

6-Ja

n-20

15

Page 50: Probabilistic Graph Models WMCI 20154

•In

other

word

s,5/8

=0.6

25or

there

are

62.5%

chan

ceso

fgett

ingco

rrect

outpu

tvolt

age.

Theo

retic

ally

•Su

bstitu

ting

•Th

us,

theor

etica

lan

dex

perim

ental

resu

ltsar

ema

tching

.

•Th

elas

tthr

eeca

ses

inthe

Table

deno

testhe

minim

alcu

tsets

forthi

snan

ocell

,den

oted

by,

•As

depic

tedfro

mTa

bleIII,

outpu

tvo

ltage

isno

tre

ceive

dfor

these

case

s. WM

SC16

-Jan

-201

5

Page 51: Probabilistic Graph Models WMCI 20154

•Fu

rther,

toev

aluate

there

liabil

ityof

atra

ined

nano

cell,

weas

sume

thatt

here

are

kred

unda

ntpa

thsfro

minp

utto

outpu

t.Ea

chof

these

paths

may

vary

inlen

gth.

The

length

ofan

ypa

thi

can

bere

pres

ented

byva

riable

•Th

atis,

each

path

cons

ists

ofl im

olecu

lesco

nnec

tedin

serie

san

dsu

chkp

athsa

rewo

rking

inpa

ralle

l.

•Fo

rcor

rect

functi

oning

ofthe

syste

m,at

least

one

ofthe

ipath

smu

stfun

ction

corre

ctly.

•Co

nside

ran

indica

torva

riable

x ijwh

ichde

notes

thesta

teof

molec

ulars

witch

jonp

athi.

WM

SC 1

6-Ja

n-20

15

Page 52: Probabilistic Graph Models WMCI 20154

•W

edefi

neas

tructu

refun

ction

φ(x) ifo

rpath

ias

•Th

enstr

uctur

efun

ction

ofthe

whole

syste

mca

nbe

given

as:

•He

nce,

there

liabil

ityr(p

)oft

hewh

olesy

stem

atan

yins

tanto

ftime

isgiv

enas WM

SC 1

6-Ja

n-20

15

Page 53: Probabilistic Graph Models WMCI 20154

•Le

t’sco

nside

rano

there

xamp

leha

ving

multip

lepa

thsfro

minp

utto

outpu

tinfig

ureb

.

•Th

enod

eAis

input

andn

ode

Fis

outpu

t

•Th

e stru

cture

func

tion

for al

l mini

mal p

aths f

rom

A to

F ar

e

WM

SC 1

6-Ja

n-20

15

Page 54: Probabilistic Graph Models WMCI 20154

•R

elia

bilit

y is

com

pute

d as

:-

WM

SC 1

6-Ja

n-20

15

Page 55: Probabilistic Graph Models WMCI 20154

Boun

ds o

n re

liabi

lity

of a

nan

ocel

l

•Le

tA1

,A2,…

……

.,As

deno

temi

nimal

path

sets

conn

ectin

ginp

utno

deto

theou

tputn

ode

andw

edefi

neFi,

i=1…

…..s

as

•By

failin

gof

themo

lecula

rcon

necti

on,w

eme

anto

sayt

hat,

molec

uleis

in‘O

FF’s

tate.

Ifat

least

one

ofthe

molec

ules

inthe

minim

alpa

thse

thas

failed

,the

syste

mwi

llfai

leve

ntuall

y.Ma

thema

ticall

y,it

isde

noted

as:

•He

ncefo

rth, it

can b

e eas

ily de

rived

that

failur

e of a

t leas

t one

mo

lecule

in m

inima

l path

Ai c

an in

creas

e the

prob

abilit

y of fa

ilure

of at

lea

st on

e mole

cule

in pa

th Aj

WM

SC 1

6-Ja

n-20

15

Page 56: Probabilistic Graph Models WMCI 20154

•Th

iswo

uldbe

theca

seifb

othpa

thsAi

andA

jove

rlap.

So-

•Si

milar

ly

•Su

bstitu

ting

ineq

uatio

nsta

tedab

ove

wege

t

WM

SC 1

6-Ja

n-20

15

Page 57: Probabilistic Graph Models WMCI 20154

•Ag

ain,l

etC1

...C

rden

otethe

minim

alcu

tsets

.We

defin

ethe

even

tsE1

,...

,Erb

y

•Si

nce,

thena

noce

llwill

functi

oniff

allof

theev

ents

Eioc

cur,w

esay

•He

nce

WM

SC 1

6-Ja

n-20

15

Page 58: Probabilistic Graph Models WMCI 20154

•So

thebo

unds

onre

liabil

ityfun

ction

given

as

•Co

nside

rthe

same

exam

plesh

own

infig

ure

a,the

relia

bility

boun

dsar

eexp

ress

edas

:

WM

SC 1

6-Ja

n-20

15

Page 59: Probabilistic Graph Models WMCI 20154

•Su

bstitu

ting

•W

eget

•Th

ese

relia

bility

boun

dsma

tchthe

value

sco

mpute

din

colum

n4

ofTa

ble.

•Th

isex

ample

can

bege

nera

lized

forna

noce

llsco

nsist

ingof

more

than

three

nano

partic

lesan

dmole

cules

ands

imila

rres

ultsc

anbe

obtai

ned.

WM

SC 1

6-Ja

n-20

15

Page 60: Probabilistic Graph Models WMCI 20154

EXPE

RIM

ENTA

L SE

TUP

FOR

RELI

ABIL

ITY

EVAL

UATI

ON

WM

SC 1

6-Ja

n-20

15

Page 61: Probabilistic Graph Models WMCI 20154

ALGO

RITH

M 3:

Nan

ocell

Relia

bility

Pre

dictio

n Algo

rithm

(NRP

A)In

put: N

:= N

umbe

r of N

anop

artic

les;

IP:=

Prim

ary I

nput

Node

;OP

:= P

rimar

y Outp

ut No

de ;

P mol

:= P

roba

bility

of a

molec

ule be

ing pr

esen

t betw

een t

wo na

nopa

rticles

and f

ound

in ‘O

N’sta

te ;

Outp

ut:

P pat

h:=

Pro

babil

ity th

at at

least

one p

ath ex

ists b

etwee

n IP

and O

P;St

eps:

1: Ini

tializ

ation

;2:

Adj_M

atrix

= Ge

nera

te M

atrix

(N,N

,µ,σ

) ;/*

Gene

rate

a NxN

rand

om a

rray (

Adjac

ency

Matr

ix) w

ith G

auss

ian D

istrib

ution

*/ ;

3: Ed

ge_M

atrix

= Co

nver

t_to

_Edg

eMat

rix (A

dj_M

atrix

) ;/*

Conv

ert th

is Ad

jacen

cy M

atrix

to Ed

ge M

atrix

*/ ;

4: Fw

d_po

inting

_ ar

ray =

Gen

erat

e Dag

(Edg

e_M

atrix

) ;/*

Remo

ve se

lf loo

ps an

d bac

kwar

d poin

ting e

dges

from

this

matrix

*/ ;

5:M

= si

ze (F

wd_p

ointin

g_ar

ray,1

) ;/*

Calcu

late N

umbe

r of M

olecu

les, d

enote

d as

’M’ *

/ ;6:

Nano

cell_

Mat

[N][N

] = W

eight

ed M

atrix

(Fwd

_poin

ting_

arra

y,Pm

ol) ;

/* Ad

d weig

ht to

each

conn

ected

edg

e = P

mol

and u

ncon

necte

d ed

ge =

∞ of

Fwd

_poin

ting_

arra

y*/ ;

7: sh

orte

st_pa

th=

Dijks

tra_k

(Nan

ocell

_Mat

,IP,O

P,k) ;

/* Ca

lculat

e ksh

ortes

t path

s fro

m IP

to O

P us

ing m

odifie

d Di

jkstra

algor

ithm

*/ ;

8: P M

= As

sign_

Prob

(Pm

ol);

/* As

sign p

roba

bilitie

s to e

ach m

olecu

le */

;9:

P pat

h=

Com

pute

_Pro

b(sh

orte

st_pa

th, P

M);

/* Ca

lculat

e pro

babil

ity P

path

, as e

xplai

ned i

n Sec

tion 4

*/ ;

retu

rn P

path

;W

MSC

16-

Jan-

2015

Page 62: Probabilistic Graph Models WMCI 20154

Sim

ulat

ion

Resu

lts

WM

SC 1

6-Ja

n-20

15

Page 63: Probabilistic Graph Models WMCI 20154

WM

SC 1

6-Ja

n-20

15

Page 64: Probabilistic Graph Models WMCI 20154

Low

er B

ound

WM

SC 1

6-Ja

n-20

15

Page 65: Probabilistic Graph Models WMCI 20154

Uppe

r Bou

nd

WM

SC 1

6-Ja

n-20

15

Page 66: Probabilistic Graph Models WMCI 20154

Dis

tribu

tion

func

tion

for n

umbe

r of

conn

ecte

d pa

ths

WM

SC 1

6-Ja

n-20

15

Page 67: Probabilistic Graph Models WMCI 20154

PROB

ABIL

ITY

ANAL

YSIS

OF

NAN

OCEL

L IN

TIM

E D

OMAI

N

WM

SC 1

6-Ja

n-20

15

Page 68: Probabilistic Graph Models WMCI 20154

Augm

ente

d Co

ntin

uous

Par

amet

er B

irth

Dea

th

Mod

el•

Cons

ider a

nano

cell h

aving

N na

nopa

rticles

conn

ected

via M

mole

cular

switc

hes.

•Le

t (t) deno

te the

numb

er of

‘ON’

mole

cules

in th

e nan

ocell

, at ti

me t

•Le

t Sj(t)

=

j(t) de

notes

the s

uper

-state

of th

e sys

tem.

•Le

t the s

tate s

pace

of th

e pro

cess

be I =

{0, 1

, 2,…

,M} a

nd T

= [0

;1) be

its pa

rame

ter

spac

e. Th

at is,

each

supe

r-stat

e den

otes n

umbe

r of ‘

ON’ m

olecu

les an

d it c

onsis

ts of

one o

r mor

e sub

-state

s. •

The t

rans

ition f

rom

a give

n stat

e to a

nothe

r stat

e can

take

plac

e at a

ny in

stant

of tim

e. •

Thus

, as s

hown

in F

ig. , a

t any

time t

, one

of th

e sub

-state

Sk j,

k = 1

to M C

jof a

supe

r-sta

te S j

, can

mak

e1)

a do

wn tr

ansit

ion to

one o

f the (

M -j)

sub-

states

of S

j+1 s

uper

-state

, with

failu

re

rate

j.2)

an up

tran

sition

to on

e of th

e j su

b-sta

tes of

Sj-1

supe

r-stat

e, wi

th re

pair r

ate

j.

WM

SC 1

6-Ja

n-20

15

Page 69: Probabilistic Graph Models WMCI 20154

Hie

rarc

hica

l CTM

C m

odel

WM

SC 1

6-Ja

n-20

15

Page 70: Probabilistic Graph Models WMCI 20154

Exam

ple

Nan

ocel

l

WM

SC 1

6-Ja

n-20

15

Page 71: Probabilistic Graph Models WMCI 20154

Algo

rithm

WM

SC 1

6-Ja

n-20

15

Page 72: Probabilistic Graph Models WMCI 20154

Sim

ulat

ion

Resu

lts

WM

SC 1

6-Ja

n-20

15

Page 73: Probabilistic Graph Models WMCI 20154

Sim

ulat

ion

Resu

lt

No. of Success per case

No. of Success per case

WM

SC 1

6-Ja

n-20

15

Page 74: Probabilistic Graph Models WMCI 20154

Prob

abili

ty o

f Sta

te R

each

abili

ty

WM

SC 1

6-Ja

n-20

15

Page 75: Probabilistic Graph Models WMCI 20154

Expe

cted

Nan

ocel

l Life

time

WM

SC 1

6-Ja

n-20

15

Page 76: Probabilistic Graph Models WMCI 20154

Sim

ulat

ion

Resu

lt

WM

SC 1

6-Ja

n-20

15

Page 77: Probabilistic Graph Models WMCI 20154

Post-

fabric

ation

Tra

ining

of N

anoc

ell M

olecu

lar M

emor

yRe

liabil

ity A

nalys

is of

Nano

cell i

n pre

senc

e of T

rans

ient E

rrors

Futur

e Wor

k

CON

CLUS

ION

S AN

D F

UTUR

E W

ORK

WM

SC 1

6-Ja

n-20

15

Page 78: Probabilistic Graph Models WMCI 20154

•Th

e Nan

ocell

base

d mole

cular

mem

ory h

as be

en m

odele

d & tr

ained

using

Omni

pote

nt T

rain

ing

•Mo

rtal T

rain

ing

•Th

e sim

ulatio

n res

ults o

f 1-b

it mole

cular

mem

ory a

re co

mpar

ed ag

ainst

analy

tical

prob

abilis

tic m

odeli

ng ap

proa

ch.

•It i

s obs

erve

d tha

t 20 o

r mor

e mole

cules

mus

t be p

rese

nt in

the na

noce

ll for

relia

ble

molec

ular m

emor

y buff

er.

•Th

e con

verg

ence

time i

s high

in pr

opos

ed m

ethod

ology

for t

raini

ng m

emor

y. So

me ne

w ad

aptiv

e meth

od is

to be

appli

ed to

redu

ce th

e tra

ining

time.

•Th

e pro

babil

istic

mode

ling a

nd an

alysis

of na

noce

ll is d

one i

n pre

senc

e of s

oft tr

ansie

nt er

rors.

An ex

tende

d Con

tinuo

us P

aram

eter B

irth D

eath

Mode

l for a

Nan

ocell

is pr

opos

ed.

•It i

s obs

erve

d tha

t the n

anoc

ell w

ill fun

ction

relia

bly as

long

as fa

ilure

rate

is les

s tha

n re

pair r

ate.

Conc

lusi

ons

and

Furth

er w

orks

WM

SC 1

6-Ja

n-20

15

Page 79: Probabilistic Graph Models WMCI 20154

•Th

e des

ign of

high

dens

ity M

olecu

lar M

emor

y is u

nder

prog

ress

.•

Reali

zatio

n of N

anoc

ell M

olecu

lar M

emor

y.•

The M

olecu

lar M

odel

need

be im

prov

ed

•Th

e effe

ct on

nano

cell d

evice

beha

vior w

hen m

ore t

han o

ne m

olecu

le ar

e pre

sent

betw

een t

wo na

nopa

rticles

. Furth

er w

orks

cont

d.

WM

SC 1

6-Ja

n-20

15

Page 80: Probabilistic Graph Models WMCI 20154

1.R.

Kuma

wat,V

.Sah

ula,a

ndM.

S.Ga

ur,“

Prob

abilis

ticmo

delin

gand

analy

sisof

molec

ular

memo

ryce

ll”,AC

MJo

urna

lon

Emer

ging

Tech

nolog

iesin

Com

putin

gSy

stem

(ACM

JETC

)2.

Renu

Kuma

wat,

Vine

etSa

hula,

Mano

jSi

ngh

Gaur

,“A

Prob

abilis

ticMo

del

forRe

liabil

ityEv

aluati

onof

Nano

cell

inpr

esen

ceof

Tran

sient

Erro

rs",(

Acce

pted

inIE

TDi

gital

&Co

mpute

r)

Publ

icat

ions

In P

eer-r

evie

wed

Jour

nals

WM

SC 1

6-Ja

n-20

15

Page 81: Probabilistic Graph Models WMCI 20154

[1]Int

erna

tiona

ltech

nolog

yroa

dmap

forse

mico

nduc

tors,.

2009

,201

0.[2]

Y.Ch

en,D

.A.A

.Jun

g,G.

and

Ohlbe

rg,X

.Li,

D.R.

Stew

art,

J.O.

Jepp

esen

,K.A

.Niel

sen,

J.Fr

aser

Stod

dart,

and

R.S.

Willi

ams,

“Nan

osca

lemo

lecula

r-swi

tchcro

ssba

rcirc

uits,”

Nano

tech

nolog

y,vo

l.14,

no.4

,pp.

462.4

68.,

2003

[3]W

.Wu,

G.-Y

.Jun

g,D.

Olyn

ick,J

.Stra

znick

y,Z.

Li,X.

Li,D.

Ohlbe

rg,Y

.Che

n,S.

-Y.W

ang,

J.Lid

dle,W

.Ton

g,an

dR.

S.W

illiam

s,“O

ne-ki

lobitc

ross

-bar

molec

ularm

emor

ycirc

uitsa

t30-

nmha

lf-pitc

hfab

ricate

dby

nano

impr

intlith

ogra

phy,.

”App

lied

Phys

icsA:

Mat

erial

sScie

nce

amp;

Proc

essin

g,vo

l.80,

pp.1

173.

1178

,200

5,10

.100

7/s0

0339

-004

-317

6-y.

[4]J.

E.Gr

een,

J.W

ookC

hoi,

A.Bo

ukai,

Y.Bu

nimov

ich,E

.Joh

nston

-Halp

erin,

E.De

Ionno

,Y.L

uo,B

.A.S

heriff

,K.X

u,Y.

Shik

Shin,

H.-R

.Tse

ng,J

.F.S

todda

rt,an

dJ.

R.He

ath,“

A16

0-kil

obit

molec

ulare

lectro

nicme

mory

patte

rned

at10

11bit

spe

rsq

uare

centi

metre

,»v

ol.44

5,pp

.414

.417,

2007

[5]J.

Tour

,W.V

anZa

ndt,

C.Hu

sban

d,S.

Husb

and,

L.W

ilson

,P.F

ranz

on,a

ndD.

Nack

ashi,

“Nan

ocell

logic

gates

formo

lecula

rco

mputi

ng,.”

Nano

tech

nolog

y,IE

EETr

ansa

ction

son,

vol.1

,no.

2,pp

.100

.109

,jun

2002

.[6]

P.Jh

aan

dV.

Sahu

la,.O

mnipo

tenta

ndmo

rtaltr

aining

ofa

nano

cellm

odel

toem

ulate

thefun

ction

ality

ofa

logic

gate,

.inIn

diaCo

nfer

ence

(INDI

CON)

,201

0An

nual

IEEE

,dec

.201

0,pp

.1.6

.[7]

M.M.

Ziegle

rand

M.R.

Stan

,.A

case

forcm

os/na

noco

-des

ign,.

InIC

CAD,

2002

,pp.

348.

352.

[8]G.

Rose

,M.Z

iegler

,and

M.St

an,.

Larg

e-sig

naltw

o-ter

mina

ldev

icemo

delfo

rnan

oelec

tronic

circu

itana

lysis,

.Ver

yLar

geSc

aleIn

tegr

ation

(VLS

I)Sy

stem

s,IE

EETr

ansa

ction

son,

vol.1

2,no

.11,

pp.1

201

.120

8,no

v.20

04.

[9]G.

Rose

,Y.Y

uxing

,J.M

.T.,

C.C.

Adam

,G.N

adine

,M.N

aban

ita,C

.B.J

ohn,

R.H.

Lloyd

,and

R.S.

Mirce

a,.D

esign

ingcm

os/m

olecu

larme

morie

swh

ileco

nside

ring

devic

epa

rame

terva

riatio

ns,.

ACM

Jour

nalo

nEm

ergin

gTe

chno

logies

inCo

mpu

ting

Syste

ms(

JETC

),vo

l.3,n

o.1,

2007

.Refe

renc

es

WM

SC 1

6-Ja

n-20

15

Page 82: Probabilistic Graph Models WMCI 20154

[10]R

.Iris

Baha

r,Jo

seph

Mund

y,an

dJie

Chen

.Apr

obab

ilistic

-bas

edde

sign

metho

dolog

yfor

nano

scale

comp

utatio

nIn

tern

ation

alCo

nfer

ence

onCo

mpu

ter-A

ided

Desig

n,pa

ges4

80–4

86,2

003.

[11]K

.Nep

al,R.

I.Ba

har,

J.Mu

ndy,

W.R

.Patt

erso

n,an

dA.

Zasla

vsky

.Des

igning

logic

circu

itsfor

prob

abilis

ticco

mputa

tion

inthe

pres

ence

ofno

ise.I

nDA

C’05

:Pro

ceed

ings

ofth

e42

ndan

nual

Desig

nAu

tom

ation

Conf

eren

ce,p

ages

485–

490,

New

York

,NY

,USA

,200

5.AC

M.

[12]

K.Ne

pal,

R.I.

Baha

r,J.

Mund

y,W

.R.P

atter

son,

and

A.Za

slavs

ky.D

esign

ingmr

fbas

eder

rorc

orre

cting

circu

itsfor

memo

ryele

ments

.In

DATE

’06:P

roce

eding

sof

the

conf

eren

ceon

Desig

n,au

tom

ation

and

test

inEu

rope

,pag

es79

2–79

3,30

01Le

uven

,Belg

ium,B

elgium

,200

6.Eu

rope

anDe

signa

ndAu

tomati

onAs

socia

tion.

[13]

K.Ne

pal,

R.I.

Baha

r,J.

Mund

y,W

.R.P

atter

son,

and

A.Za

slavs

ky.D

esign

ingna

nosc

alelog

iccir

cuits

base

don

marko

vra

ndom

fields

.J.E

lectro

n.Te

st.,2

3(2-

3):2

55–2

66,2

007.

[14]D

ong

S.Ha

,Bino

yRav

indra

nMe

mber

,Deb

ayan

Bhad

uri,a

ndDe

baya

nBh

adur

i.Too

lsan

dtec

hniqu

esfor

evalu

ating

relia

bility

trade

-ffos

forna

no-a

rchite

cture

s.In

inNa

no,Q

uant

uman

dM

olecu

larCo

mpu

ting:

Impli

catio

nsto

High

Leve

lDes

ignan

dVa

lidat

ion.K

luwer

Acad

emic

Publi

sher

s,20

04.

[15]D

ebay

anBh

adur

iand

Sand

eep

Shuk

la.Na

nolab

:Atoo

lfore

valua

ting

relia

bility

ofde

fect-t

olera

ntna

noar

chite

cture

s.In

inIE

EECo

mpu

terS

ociet

yAnn

ualS

ympo

sium

onVL

SI(IE

EE,p

age

0309

.Pre

ss,2

004.

[16]

Mano

jS.G

aur,

Ragh

aven

dra

Nara

simha

n,Vi

jayLa

xmi,

and

Ujjw

alKu

mar.

Stru

ctura

lfau

ltmo

dellin

gin

nano

devic

es.I

nNa

noNe

t,pa

ges6

–10,

2008

.[17

]S.

Bhan

jaan

dN.

Rang

anath

an.S

witch

ingac

tivity

estim

ation

ofvls

icirc

uits

using

baye

sian

netw

orks

.Ver

yLa

rge

Scale

Inte

grat

ion(V

LSI)

Syste

ms,

IEEE

Tran

sacti

onso

n,11

(4):5

58–

567,

aug.

2003

.[18

]San

jukta

Bhan

jaan

dN.

Rang

anath

an.C

asca

ded

baye

sian

infer

encin

gfor

switc

hing

activ

ityes

timati

onwi

thco

rrelat

edinp

uts.

IEEE

Tran

s.VL

SISy

st.,1

2(12

):136

0–13

70,2

004.

[19]T

hara

Rejim

onan

dSa

njukta

Bhan

ja.An

accu

rate

prob

alisti

cmod

elfor

erro

rdete

ction

.In

VLSI

D’05

:Pr

ocee

dings

ofth

e18

thIn

tern

ation

alCo

nfer

ence

onVL

SIDe

sign

held

joint

lywi

th4t

hIn

tern

ation

alCo

nfer

ence

onEm

bedd

edSy

stem

sDe

sign,

page

s717

–722

,Was

hingt

on,D

C,US

A,20

05.I

EEE

Comp

uterS

ociet

yW

MSC

16-

Jan-

2015

Page 83: Probabilistic Graph Models WMCI 20154

[20]

T.Re

jimon

and

S.Bh

anja.

Scala

blepr

obab

ilistic

comp

uting

mode

lsus

ingba

yesia

nne

twor

ks.I

nCi

rcuit

san

dSy

stem

s,20

05.4

8th

Midw

estS

ympo

sium

on,p

ages

712

–715

Vol.1

,7-1

020

05.

[21]

Karth

ikeya

nLin

gasu

bram

anian

and

Sanju

ktaBh

anja.

Prob

abilis

ticma

ximum

erro

rmo

delin

gfor

unre

liable

logic

circu

its.I

nGL

SVLS

I’07

:Pro

ceed

ings

ofth

e17

thAC

MGr

eatL

akes

sym

posiu

mon

VLSI

,pag

es22

3–22

6,Ne

wYo

rk,NY

,USA

,200

7.AC

M.[22

]T.

Rejim

on,K

.Ling

asub

rama

nian,

and

S.Bh

anja.

Prob

abilis

ticer

rorm

odeli

ngfor

nano

-dom

ainlog

iccir

cuits

.Ver

yLa

rge

Scale

Inte

grat

ion(V

LSI)

Syste

ms,

IEEE

Tran

sacti

onso

n,17

(1):5

5–

65,ja

n.20

09.

[23]

A.Ab

dolla

hi.Pr

obab

ilistic

decis

iondia

gram

sfor

exac

tpro

babil

istic

analy

sis.I

nCom

pute

r-Aide

dDe

sign,

2007

.ICC

AD20

07.I

EEE/

ACM

Inte

rnat

ional

Conf

eren

ceon

,pag

es26

6–2

72,2

007.

[24]

Smita

Krish

nasw

amy,

Geor

geF.

Viam

ontes

,Igo

rL.M

arko

v,an

dJo

hnP.

Haye

s.Ac

cura

tere

liabil

ityev

aluati

onan

den

hanc

emen

tvia

prob

abilis

tictra

nsfer

matric

es.I

nDe

sign,

Auto

mat

ionan

dTe

stin

Euro

peCo

nfer

ence

and

Expo

sition

(DAT

E20

05),

7-11

Mar

ch20

05,M

unich

,Ger

man

y,pa

ges2

82–2

87.I

EEE

Comp

uterS

ociet

y,20

05.

[25]

Denis

Teixe

iraFr

anco

,Mai

Corre

iaVa

scon

celos

,Lirid

Navin

er,a

ndJe

an-F

ranc

oisNa

viner

.Sign

alpr

obab

ility

forre

liabil

ityev

aluati

onof

logic

circu

its.M

icroe

lectro

nicsr

eliab

ility,

48(8

-9):1

586

–15

91,2

008.

[26]

H.De

rinan

dP.

A.Ke

lly.D

iscre

te-ind

exma

rkov-t

ype

rand

ompr

oces

s.In

proc

eedin

gsof

the

IEEE

,volu

me

77,p

ages

1485

–151

0,oc

t.19

89.

[27]

Julia

nBe

sag.

Spati

alint

erac

tion

and

thesta

tistic

alan

alysis

oflat

tice

syste

ms.J

ourn

alof

the

Roya

lSta

tistic

alSo

ciety.

Serie

sB(M

etho

dolog

ical),

36(2

):192

–236

,197

4.[28

]Jo

natha

nS

Yedid

ia,W

illiam

TFr

eema

n,an

dYa

irW

eiss.

Unde

rstan

ding

belie

fpro

paga

tion

and

itgen

erali

zatio

ns.I

nEx

plorin

gar

tificia

lint

ellige

nce

inth

ene

wm

illenn

ium,

page

s23

9–26

9,Sa

nFr

ancis

co,

CA,

USA,

2003

.Mo

rgan

Kaufm

ann

Publi

sher

sInc

.[29

]R.

Kuma

wat,

V.S.

Sahu

la,an

dM.

S.Ga

ur.R

eliab

lecir

cuit

analy

sisan

dde

sign

using

nano

scale

devic

es.I

nIn

tern

ation

alCo

nfer

ence

onCo

mm

unica

tion

and

Elec

tronic

sSy

stem

Desig

n(IC

CESD

2013

),SP

IEpr

ocee

eding

son

,vo

lume

8760

,pag

es87

602C

–1–

8760

2C–8

,jan

2013

.[30

]Ju

dea

Pear

l.Pr

obab

ilistic

Reas

oning

inIn

tellig

entS

yste

ms:

Netw

orks

ofPl

ausib

leIn

fere

nce.

Mor

gan

Kauf

man

nPu

blish

ersI

nc.,

1988

.

WM

SC 1

6-Ja

n-20

15

Page 84: Probabilistic Graph Models WMCI 20154

1.“In

trodu

ction

to pr

obab

ility m

odels

” by S

heldo

n M. R

oss

2.“A

pplie

d Pro

babil

ity M

odels

with

Opti

miza

tion A

pplic

ation

s” by

She

ldon M

. Ros

s3.

“Pro

babil

ity M

odels

for C

ompu

ter S

cienc

e” by

She

ldon M

. Ros

s4.

“Sim

ulatio

n” by

She

ldon M

. Ros

s5.

“Gra

ph T

heor

y with

App

licati

ons t

o Eng

ineer

ing an

d Com

puter

Scie

nce”

by N

arsin

ghDe

o6.

“Pro

babil

istic

Reas

oning

in In

tellig

ent S

ystem

s ” b

y J. P

earl

7.“P

roba

bility

and S

tatist

ics w

ith R

eliab

ility ,

Queu

ing an

d Com

puter

Scie

nce

Appli

catio

ns” b

y Kish

orS.

Triv

edi

Book

s

WM

SC 1

6-Ja

n-20

15

Page 85: Probabilistic Graph Models WMCI 20154

WM

SC 1

6-Ja

n-20

15

Page 86: Probabilistic Graph Models WMCI 20154

WM

SC 1

6-Ja

n-20

15

Page 87: Probabilistic Graph Models WMCI 20154

WM

SC 1

6-Ja

n-20

15

Page 88: Probabilistic Graph Models WMCI 20154

WM

SC 1

6-Ja

n-20

15

Page 89: Probabilistic Graph Models WMCI 20154

WM

SC 1

6-Ja

n-20

15

Page 90: Probabilistic Graph Models WMCI 20154

WM

SC 1

6-Ja

n-20

15