Neural Mechanisms of Object Perception

appliancepartΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 4 χρόνια και 21 μέρες)

124 εμφανίσεις

Neural Mechanisms of Object Perception

Zhiyong Yang



Brain and Behavior Discovery Institute

James and Jean Culver Vision
Discovery Institute

Department of Ophthalmology

Georgia Regents University


April

4, 2013

Outline

1.
A

model

of

pattern

recognition


2
.

An

updated

view

of

the

ventral

pathway


3
.

Neural

codes

for

object

perception




3
.
1
.

V
1

and

V
2



3
.
2
.

V
4


3
.
3
.

IT

4
.

A

perspective

based

on

untangling

object

manifolds





A
Model
of
Pattern Recognition


Features



Probability

distributions


Decision

rule




The Ventral Pathway

Kravitz et. al., 2013

Occipitotemporal
network

Output pathways

Three

Cortico
-
subcortical
Output
Pathways

1.

Occipitotemporo
-
neostriatal

pathway



reinforcement

learning


2
.

Occipitotemporo
-
ventral

striatum

pathway




value


3
.

Occipitotemporo
-
amygdaloid

pathway




emotion







Three Cortico
-
cortical

Output
Pathways

1
.

Occipitotemporo
-
medial

temporal

pathway



long
-
term

memory

2
.

Occipitotemporo
-
orbitofrontal

pathway




reward

3
.

Occipitotemporo
-
ventrolateral

pathway




working

memory

and

executive

function







Neural codes for object perception

Neural Codes in V1

1.
Responses

selectively

to

a

full

range

of

visual

features


orientation,

direction,

disparity,

speed,

luminance,




contrast,

color,

and

spatial

frequency

2
.

Functional

maps



retinotopic

map,

orientation

map,

ocular

dominance



map

3
.

Contextual

modulation

4
.

Adaptive

5
.

Sparse

and

decorrelated

relative

to

inputs


Orientation Selectivity

Hubel & Wiesel, 1968

Orientation Map

Nauhaus et. al., 2008

LNL Models of V1 Neurons

simple cells

complex cells

Shape Codes in V2

1.
Responses

to

single

orientation

2
.

Responses

to

multiple

orientations

3
.

Responses

to

shapes

of

intermediate



complexity

Anzai et. al. 2007

Stimulus sets

Grating

stimuli

Contour

stimuli

Hegde & Van Essen, 2007

Response profiles of exemplar V4 and V2 cells

Shape Codes in V4

1.
Responses

selectively

to

curvature,

orientation,

and

object
-
relative

position


2
.

Evidence

for

a

sparse

coding

scheme


Pasupathy & Connor 2002

Carlson et. al., 2011

S
parseness Index = 0.80

S
parseness Index = 0.36

S
parseness Index = 0.22

S
parseness Index = 0.11

Neural Codes in IT

1.

Structural,

configurational,

and


compositional

for

both

2
D

and

3
D

objects

2
.

Position,

orientation,

curvature

3
.

Skeletal

shape

and

boundary

shape

3
.

Structural

and

holistic

4
.

Categorical

clustering


Brincat & Connor, 2004

2D contour shapes

Brincat & Connor, 2004

2D contour shapes

Yamane
et. al., 2008

3D shapes

Yamane
et. al., 2008

3D shapes

Categorical Coding

Kriegeskorte
et. al., 2008

A perspective based on
untangling

object
manifolds

1.
Core object recognition and IT codes

2.
Untangling
object
manifolds and a
proposal

3.
Open questions


DiCarlo et. al., 2012

Core Object
Recognition

1.
Discriminate

a


visual

object

from

all

other

possible

visual

objects

within

<
200

ms
.

2.
Discount

changes

due

to

changes

in

illumination,

object

position,

size,

scale,

viewpoint
,

and

visual

context,

and

other

structural

variations
.

3.
Comprise

between

invariance

and

generalization
.


4.

There

are

~
30
,
000

natural

objects
.

5.

Current

models

approach

at

best

~
5
%

of

human

object

perception
.


Untangling
Object Representations

The
Ventral Visual
Pathway


Each
area
proportional
to
cortical
surface
area.

Total
number of
neurons. Dimensionality
of each
representation

Portion
(color) dedicated
to

processing
the central 10 deg of
the visual field

Median
response latency

IT Neural Codes

1.
Spike

counts

in

~
50

ms

convey

information

object

identity

2
.

Object

identity

information

is

available

~
100

ms

after



presentation

3
.

IT

population

presentation

is

untangled

and

object



identity

can

be

decoded

by

weighted

summation

codes
.

4
.

These

codes

are

quite

general
.

IT Single
-
Unit Properties and Their Relationship to
Population
Performance

Abstraction Layers and Their Potential
Links

Serial
-
Chain Discriminative

Models of Object
Recognition

A Neural Network Model of Object
Recognition

Serre et. al., 2007


A Model of Object Recognition