[PPT] - Explaining Cortical Adaptation with a Statistically Optimized PowerPoint Presentation

SLIDE 1 Explaining Cortical Adaptation with a Statistically Optimized Normalization Mo del Martin W ain wrigh t Eero Simoncelli Vision Sciences Cen ter for Neural Science Harv ard Univ ersit y Couran t Institute New Y

rk

Univ ersit y

SLIDE 2 In tro duction Hyp

thesis

sensory systems are matc hed to their input statistics A ttnea v e

More

sp ecically statistical indep endence

f

neural resp

nses

Barlo w

Role
f

image statistics

indep

endence

f

resp

nses

m ust b e dened with resp ect to statistics

f

visual input

large

b

dy
f

previous researc h

n

natural image statistics and cortical pro cessing eg Field

A

tic k

Redlic

h

v

an Hateren

Ruderman
Olshausen
Field
Bell
Sejno

wski

SLIDE 3 Cortical adaptation and image statistics Statistics

f

visual input are constan tly c hanging

v

er seconds andor min utes Question Can cortical adaptation b e understo

d

as

ptimal

adjustmen t to statistics

f

recen t input Sev eral authors ha v e tried to link input statistics to cortical adaptation eg Barlo w

W

ain wrigh t

Limitations
f

previous w

rk
simplistic

mo dels

f

images eg Gaussian

linear

mo dels

f

neurons

SLIDE 4 Normalization mo dels

Divisiv

e normalization

Compute

linear resp

nses

fL k g

f

receptiv e elds at dier en t spatial scales p

sitions

and

rien

tations

Compute

a normalized resp

nse

b y dividing a cells squared resp

nse

L

b

y a sum

f

squared re sp

nses
f

neigh b

rs
Normalization

accoun ts for nonlinear b eha vior in neurons Bonds

Geisler

and Albrec h t

Heeger
Normalization

can b e deriv ed from natural image statistics Simoncelli

Simoncelli

and Sc h w artz

SLIDE 5 Statistical view

f

normalization

normalization

is a form

f

nonline ar pr e dictive c

ding
resp
nses
f

neigh b

ring

mo del neurons are used to pr e dict the v ariance

f

a mo del neuron

mo

del neuron is normalized b y the prediction R

L
P

k

k

jL k j

normalized

resp

nses

are close to statistically indep enden t Key P

in

t

and

f k g are determined b y the statistics

f

the visual en vironmen t

SLIDE 6 Con trast adaptation Increase con trast

increase
shift

CRF righ t

Environment A −200 200 −200 200

R A

L
A

L

Environment B

−400 400 −400 400

R B

L
B

L

1

10 100 10

−2

10

−1

10 Rightward shift Log response % Grating contrast

Environment A Environment B

SLIDE 7 P attern adaptation Increase dep endency

increase
decrease

saturation

Environment A −200 −100 100 200 −200 −100 100 200

R A

L
A

L

Environment B

−200 −100 100 200 −200 −100 100 200

R B

L
B

L

1

10 100 10

−2

10

−1

10 Saturation change Log response % Grating contrast

Environment A Environment B

SLIDE 8 Sim ulation

f

adaptation

Compute

generic parameters for an en vironmen t

f

natural images

Compute

adapte d parameters for a mixture

f

sine w a v e grating and natural images

Compute

normalized resp

nses

to sin usoidal test stim uli using eac h set

f

parameters

SLIDE 9 CRF Dieren t adapting con trasts Cell Mo del Albrec h t et al

1

10 100 1 10 100 Contrast (%) Response (spikes/s)

Low contrast adapt High contrast adapt

1 10 100 1 10 100 Contrast (%) Response

High contrast adapt Low contrast adapt

SLIDE 10 CRF Dieren t test spatial frequencies Cell Mo del Albrec h t et al

2

10 50 1 10 100 Contrast (%) Response (spike/s) Optimal Test 2 10 50 1 10 100 Contrast (%) Non−optimal Test

Unadapted Adapted

2 10 50 1 10 100 Contrast (%) Response Optimal Test 2 10 50 1 10 100 Contrast (%) Non−optimal Test

Adapted Unadapted

SLIDE 11 T uning curv es Dieren t adapting

rien

tations Cell Mo del M

uller
Lennie
−40

−20 20 40 0.2 0.4 0.6 0.8 1

Orientation (deg) Response

Unadapted Adapt 14 Adapt 0

−60 −30 30 60 0.2 0.4 0.6 0.8 1

Orientation (deg) Response

Adapt 14 Unadapted Adapt 0

SLIDE 12 Conclusions

Cortical

adaptation can b e explained using a normalization mo del with parameters determined b y image statistics

Suc

h a mo del mak es a principled distinction b et w een con trast and pattern adaptation

Mo

del accoun ts for V cell b eha vior under a v ariet y

f

adaptation conditions

SLIDE 13 References A tic k J and Redlic h A

What

do es the retina kno w ab

ut

natural scenes Neur al Computation

A

ttnea v e F

Informational

asp ects

f

visual pro cessing Psycho lo gic al R eview

Barlo

w H

A

theory ab

ut

the functional role and synaptic mec h anism

f

visual aftereects In Blak emore C editor VisionCo ding and Eciency Cam bridge Univ ersit y Press Barlo w H B

P
ssible

principles underlying the transformation

f

sensory messages In Rosen blith W A editor Sensory Commu nic ation page

MIT

Press Cam bridge MA Bell A J and Sejno wski T J

The

indep enden t comp

nen

ts

f

natural scenes are edge lters Vision R ese ar ch

Bonds

A B

Role
f

inhibition in the sp ecication

f
rien

tation

f

cells in the cat striate cortex Visual Neur

scienc

e

SLIDE 14 Field D

Relations

b et w een the statistics

f

natural images and the resp

nse

prop erties

f

cortical cells Journal

f

the Optic al So ciety

f

A meric a A Geisler W S and Albrec h t D G

Cortical

neurons Isolation

f

con trast gain con trol Vision R ese ar ch

Heeger

D J

Normalization
f

cell resp

nses

in cat striate cortex Visual Neur

scienc

e

Olshausen

B and Field D

Natural

image statistics and ecien t co ding Network Computation in Neur al Systems

Ruderman

D L and Bialek W

Statistics
f

natural images Scaling in the w

ds

Phys R ev L etters

Simoncelli

E P

Statistical

mo dels for images Compression restoration and syn thesis In st Asilomar Conf Signals Systems and Computers pages

P

acic Gro v e CA IEEE Sig Pro c So ciet y

SLIDE 15 Simoncelli E P

and

Sc h w artz O

Image

statistics and corti cal normalization mo dels In Neur al Information Pr

c

essing Systems v

lume
W

ain wrigh t M J

Visual

adaptation as

ptimal

information transmission Vision R ese ar ch