Expertise in object recognition - Department of Psychology

brasscoffeeAI and Robotics

Nov 17, 2013 (3 years and 6 months ago)

55 views

Further Explorations
of Expert Object
Recognition

Assaf Harel

Department of Psychology,

Hebrew University, Jerusalem


Acknowledgements


Prof. Shlomo Bentin


Prof. Rafael Malach


Yulia Golland

What is Expert Object
Recognition?

Experts

have

more

experience

with

and

are


more

knowledgeable

about

objects

in

their

domain

of

expertise
.


Theoretical

question
:

What

role

does


experience play in the object recognition

system?


1)

Cognitive substrate

2)

Neural substrate


Expertise: a downward shift to
subordinate level.


But, what type of processing is
needed for subordinate level
(expert) identification?



Are

faces

special?



Object

recognition

is

domain

general
.

Face

and

object

processing

should

not

be

functionally

independent
.

Faces

are

an

example

of

stimuli

which

observers

have

gained

natural

expertise

with
.




Therefore,

according

to

this

hypothesis

experts

in

any

kind

of

object

recognition

will

show

“face
-
specific”

effects
.



And Then Came Faces…

The Face Expertise Model



Neuroimaging Findings


Gauthier

et

al
.

(
2000
)

suggested

that

FFA,

a

region

which

shows

preferential

activation

for

faces,

can

also

be

activated

while

bird

and

car

experts

viewed

objects

in

their

domain

of

expertise
.




FFA

activation

was

also

found

for

laboratory

created

expertise

with

Greebles

(Gauthier

et

al
.
,

1999
)
.


Focus of Present Research


Expert object recognition in and of itself,
independent of the domain specificity/expertise
debate.



Research question: How is expert object
recognition expressed in the brain?


Are there any other “expertise” areas, except for the
FFA?


How early in the visual stream can we find expertise
effects?


Expertise
-
specific areas or a network of expertise?


Selection of Car experts


14

car

experts

(all

males,

mean

age

27
)

were

selected

based

on

their

performance

on

a

car

discrimination

task
.



Participants

had

to

determine

whether

two

cars

were

of

the

same

model

(within

maker)

varying

in

year,

color,

and

orientation
.

Their

accuracy

(d’)

on

this

task

was

1
.
39

compared

with

a

group

of

20

novices

whose

accuracy

was

0
.
57

(t(
32
)=
7
.
72
,

p<
0
.
01
)
.



In

a

similar

recognition

task,

with

a

different

object

category

(airplanes),

experts

were

as

accurate

as

novices

(
0
.
67

and

0
.
43
,

respectively,

t(
32
)=
1
.
72
,

p=
0
.
09
)
.

Methods


Participants: 9 car experts (males, mean age 23)
and 10 novices (males, mean age 25).



1
.
5
-
T

Signa

Horizon

LX

8
.
25

GE

scanner

of

the

Tel

Aviv

Sourasky

Medical

Center
.



fMRI

parameters
:

Gradient

EPI

sequence

(TR=
3000

ms,

TE=
55

ms,

FOV=
240
x
240

mm,

matrix

size=

80
x
80
,

slice

thickness=
4

mm,

1

mm

gap,

27

axial

slices)
.



T
1
-
weighted

high

resolution

(
1
x
1
x
1

mm)

anatomic

images

and

a

whole
-
brain

SPGR

sequence
.



fMRI Experiment


Three scans:


2
experimental scans (faces, cars,
airplanes)


One
-
back memory task



External localizer (faces, houses,
tools, geometric patterns)


One
-
back covert memory task


All stimuli were

equiluminant

Experimental

Design

Experts

Novices

Faces > Houses

Results: Group Activation
Maps

Novices

Cars > Airplanes

Experts

Cars > Airplanes

Experts

Novices

Cars > Airplanes


FFA

was

activated

by

faces

in

experts

and

novices
.

FFA

was

not

preferentially

activated

in

car

experts

while

viewing

cars
.



In

contrast

to

faces,

the

pattern

of

activation

elicited

by

cars

was

different

for

experts

and

novices
.



Whereas

in

novices,

activation

was

limited

to

medio
-
occipital

regions,

in

experts

the

car
-
activation

was

wide
-
spread,

distributed

over

a

large

portion

of

the

occipital

cortex

and

extending

to

posterior

regions

of

the

inferior

temporal

lobe
.




ROI Analysis

Four ROIs were defined using the external
localizer:


1.
FFA


defined by the contrast Faces vs.
Houses.

2.
LO
-

defined by the contrast Tools vs.
Textures.

3.
CoS
-

defined by the contrast Houses vs.
Faces.

4.
Early visual areas
-

defined by the contrast
Textures vs. all other object categories.

LO
0
0.2
0.4
0.6
0.8
1
1.2
experts
novices
mean % signal change
Cars
Faces
Airplanes
FFA
0
0.2
0.4
0.6
0.8
1
1.2
experts
novices
mean % signal change
Cars
Faces
Airplanes
ROI Analysis: Results

*

*

CoS
0
0.2
0.4
0.6
0.8
1
1.2
experts
novices
mean % signal change
Cars
Faces
Airplanes
Early Visual Areas
0
0.2
0.4
0.6
0.8
1
1.2
experts
novices
mean % signal change
Cars
Faces
Airplanes
*

*


The

analysis

of

the

pre
-
defined

ROIs

revealed

no

difference

between

car
-
activation

in

experts

and

novices

neither

in

the

FFA

nor

in

the

LO
.



In

early

visual

areas,

equivalent

activation

was

found

across

categories

in

novices,

while

in

experts

cars

elicited

significantly

higher

activation

than

faces

and

airplanes
.

A

similar

trend

was

found

in

the

CoS
.




What Can We Learn About
Expert Object Recognition?


The

neural

substrates

of

car

expertise

are

not

equivalent

to

those

of

face

expertise
.



Expert

object

recognition

is

distributed

and

not

restricted

to

a

specific

“hot

spot”

such

as

the

FFA
.


Expert

object

recognition

in

different

domains

recruits

different

brain

regions
.

Is

face

recognition

the

right

model

for

expert

object

recognition?





Alternative

explanation
:

the

extent

of

activity

for

objects

of

expertise

(such

as

cars),

which

is

not

seen

for

faces,

might

indicate

general

alertness/arousal

(emotional

reaction?)

superimposed

on

peculiar

perceptual

processes
.




Mourao
-
Miranda et al.
(Neuroimage,
2003
)



The

results

suggest

a

notion

of

a

dedicated

expert

object

recognition

network,

whereby

early

vision

is

top
-
down

modulated

according

to

differential

recognition

goals
.



Theoretical

framework
:

Reverse

Hierarchy

Theory

(Ahissar

&

Hochstein,

TiCS

2004
)

What Can We Learn About
Expert Object Recognition?

Hochstein & Ahissar, 2002

Ahissar & Hochstein, 2004

Basic Level Categorization

(generalization)


Subordinate
Level
Categorization
(specificity)


In

situations

requiring

better

signal
-
to
-
noise

ratio

(such

as

discriminating

among

similar

members

of

the

same

class)

highly
-
trained

performers,

who

have

had

a

great

deal

of

training

experience,

base

their

performance

on

low
-
level

representations

guided

by

top
-
down

activated

pathways
.



In

the

present

study,

this

model

manifests

in

the

car

selective

activation

found

in

early

visual’

areas

of

car

experts

and

in

the

extensive

pattern

of

activation

for

objects

of

expertise
.



Future Research


Neuroimaging studies manipulating
alertness/interest/arousal.



Behavioral measures manipulating low
-

level processing.



Temporal dynamics of low
-
level
processing vs. high
-
level processing in
experts.

RHT and Perceptual Learning

Perceptual Learning

= practice
-
induced improvement
in the ability to perform specific perceptual tasks.


1.
Perceptual learning improvement largely stems
from a gradual top
-
down guided increase in
usability in first high and then lower
-
level task
-
relevant information.


2.
This process is subserved by a cascade of top
-
to
-
bottom level modifications that enhance task
-
relevant, and prune irrelevant, information.