Sensitive, specific, and generative face recognition in a newborn visual system

PDF

Download

Open document

Flip pages

Copy asset link

Request this asset

Request accessible transcript

Transcript (if available)

Content Running
Head:
FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM
1

Sensitive,
specific,
and
generative
face
recognition
in
a
newborn
visual

system

Samantha
M.
Waters

(In
collaboration
with
Justin
N.
Wood)

University
of
Southern
California

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

2

TABLE
OF
CONTENTS

Abstract
.......................................................................................................................................................................
3

Introduction
&
Background
................................................................................................................................
4

Results
..........................................................................................................................................................................
9

Discussion
................................................................................................................................................................
13

Methods
&
Materials
...........................................................................................................................................
16

References
...............................................................................................................................................................
19

Supplementary
Information
............................................................................................................................
21

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

3

Abstract

Despite
the
computational
difficulty,
adults
easily
recognize
faces
with
incredibly
high

sensitivity
and
specificity
across
novel
changes
in
viewing
conditions.
Mature
face

representations
are
invariant
to
identity-‐preserving
changes
(e.g.,
changes
in
expression
or

viewing
angle)
while
intolerant
to
identity-‐transforming
changes
(e.g.,
changes
in
texture,

age,
and
gender).
Yet,
due
to
constraints
in
testing
newborn
subjects,
the
origins
of
these

characteristics
remain
unknown.
Are
highly
sensitive,
specific,
and
generative
face

representations
the
product
of
extensive
visual
experience,
or
do
genes
build
neural

machinery
that
generates
sensitive,
specific,
and
generative
face
representations
from
the

onset
of
experience
with
faces?
To
probe
the
nature
of
the
first
face
representation
created

in
the
newborn
mind,
we
reared
newborn
chickens
for
one
week
in
a
virtual
reality

chamber
that
contained
no
objects
except
a
single
virtual
human
face
(the
“Virtual
Parent”).

Immediately
after,
the
chickens
were
tested
for
one
week
on
their
ability
to
discriminate

between
the
Virtual
Parent
and
a
set
of
distractor
faces.
Distractors
were
identical
to
the

Virtual
Parent,
with
the
exception
of
a
single
dimension
(e.g.,
changes
in
texture,
structure,

inversion,
and
expressions).
Subjects
were
able
to
distinguish
the
Virtual
Parent
from

distractors
that
differed
in
texture,
age,
gender,
and
inversion,
but
not
distractors
that

differed
in
feature
configuration
or
facial
expression.
Further,
subjects
were
able
to

recognize
the
Virtual
Parent
across
novel
changes
in
head
orientation
(i.e.,
view
direction),

indicating
that
subjects
built
invariant
face
representations.
These
results
show
that
the

first
face
representation
built
by
the
newborn
visual
system
can
be
sensitive,
specific,
and

generative.

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

4

Sensitive,
specific,
and
generative
face
recognition
in
a
newborn
visual
system

Despite
the
computational
difficulty,
the
adult
visual
system
recognizes
faces
in

novel
viewing
conditions
with
impressive
sensitivity
(identifying
positive
matches)
and

specificity
(identifying
mismatches).
Face
recognition
has
critical
evolutionary
importance

due
to
its
role
in
social
cognition
and
successful
social
interaction.
Indeed,
newborn
infants

(1)
and
chickens
(2)
show
a
preference
for
attending
to
face-‐like
stimuli
from
birth,

suggesting
innate
mechanisms
that
orient
attention
to
specific
structural
configurations
or

‘sensory’
features
(e.g.,
dark/light
vertical
asymmetry,
horizontal
symmetry,
and
high

contrast)
(3).
A
plethora
of
research
indicates
that
our
brains
process
faces
differently
from

other
images,
using
different
cognitive
strategies
and
neural
regions
than
other
types
of

image
processing
(4-‐8).

However,
due
to
the
challenges
of
testing
newborns,
there
is
still
significant
debate

over
what
machinery
for
face
recognition
is
inherent
to
the
newborn
mind
and
how
that

machinery
is
shaped
by
experience.
Evidence
from
patients
who
had
cataracts
as
children

(but
corrected
vision
in
later
life)
suggests
that
many
elements
of
face
processing
are

learned
through
experience,
and
subject
to
a
sensitive
developmental
period
(9-‐11)
.

Absence
of
visual
experience
during
this
sensitive
period
disrupts
some
natural
processing

abilities
(e.g.,
the
composite
face
effect
and
the
ability
to
distinguish
faces
based
on
featural

spacing).
However,
these
patients
are
still
able
to
recognize
faces
based
on
features
and

contour
information
(10).

While
research
on
cataract
patients
suggests
that
visual
experiences
are
critical
to

normal
face
perception,
other
studies
have
pointed
to
face
recognition
abilities
in
subjects

with
highly
constrained
visual
experience.
Neonates
who
are
only
three
days
old

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

5

discriminate
their
mother’s
face
from
a
stranger’s
face
(12).
In
an
elegant
controlled

rearing
study,
Sugita
(13)
raised
infant
macaques
without
any
exposure
to
faces
for
6-‐24

months.
The
macaques
showed
a
preference
for
viewing
faces
rather
than
objects,
and

were
able
to
discriminate
faces
that
varied
by
features
and
featural
spacing.
These
findings

suggest
that
face
recognition
may
rely
on
innate
building
blocks
of
face
processing.

However,
the
macaques
in
Sugita’s
study
had
extensive
experience
with
visual
objects.
This

raises
the
possibility
that
the
face
recognition
abilities
seen
in
Sugita’s
study
emerged
from

domain-‐general
processes
that
were
tuned
by
subjects’
extensive
object
experiences.
The

visual
system
is
highly
plastic
and
uses
statistical
redundancies
present
in
the
natural

world
to
fine-‐tune
the
response
of
neurons
(14-‐16),
so
the
macaques’
robust
experience

with
objects
may
have
informed
their
perception
of
faces.
The
goal
of
the
present
study
was

to
test
what
types
of
face
representations
are
possible
to
build
in
the
newborn
mind

without
prior
visual
experiences
with
faces
or
objects.

Past
studies
of
the
newborn
mind
have
faced
significant
challenges.
Human
infants

cannot
ethically
be
raised
in
controlled
environments
from
birth,
and
collecting
more
than

a
few
trials
from
each
subject
is
typically
impossible.
These
difficulties
have
limited

researchers’
abilities
to
study
the
mind’s
inherent
machinery
and
how
that
machinery
is

molded
over
time
through
experience.
To
probe
the
first
face
representation
created
by
the

newborn
brain,
we
used
a
new
controlled
rearing
method
in
which
an
animal
model—the

domestic
chicken—can
be
raised
from
birth
for
several
weeks
entirely
within
a
virtual

reality
chamber.
These
chambers
provide
complete
control
over
all
visual
object
and
face

experiences
from
birth.
Face
stimuli
were
presented
to
the
subject
by
projecting
animated

videos
onto
two
virtual
walls
situated
on
opposite
sides
of
the
chamber
(Figure
1A
&
B).

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

6

The
chambers
contained
extended
surfaces
only.

Food
and
water
were
provided
in

recessed
compartments,
and
the
subjects
were
fed
“sand-‐like”
grain
because
grain
does
not

have
the
properties
of
a
solid
objects
with
a
bounded
shape.
Thus,
the
chambers
contained

no
objects
other
than
the
virtual
face
presented
on
the
virtual
walls.
The
chicken
is
an
ideal

animal
model
for
studying
the
origins
of
face
recognition.
Like
humans,
chickens
build

object
representations
and
can
recognize
2-‐D
and
3-‐D
shapes
(17-‐19).
They
also
have

neural
substrates
that
are
similar
to
the
mammalian
neocortex
(20).
However,
unlike

humans,
chickens
are
a
precocial
species
with
early
motor
development,
allowing
them
to

explore
their
environment
immediately
after
birth.
Chickens
do
not
require
parental
care,

allowing
us
to
raise
them
in
an
environment
completely
devoid
of
objects
and
non-‐virtual

companions.

Figure
1

MONITOR 1!
MONITOR 2!
Input Phase (Week 1):!
M1! M2!
…! …!
!"#$%&'("($)%*+,
'%
!" #"
$"
FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

7

A)
Chickens
were
raised
from
birth
for
one
week
in
a
world
containing
no
objects
other

than
a
single
virtual
face.
B)
The
Input
Phase
presented
the
Virtual
Parent
on
alternating

virtual
walls
(Monitor
1
and
Monitor
2).
C)
All
subjects
were
imprinted
to
the
same
Virtual

Parent.
During
the
Input
Phase
(one
week),
the
Virtual
Parent
was
the
only
stimulus

presented
to
the
subjects
on
the
virtual
walls.
In
Experiment
1,
the
Virtual
Parent’s
face

rotated
through
180°;
while
in
Experiment
2,
the
Virtual
Parent’s
face
rotated
through
10°.

Additionally,
our
design
takes
advantage
of
filial
imprinting,
a
rapid
learning
process

in
which
a
chicken
develops
a
social
preference
for
a
conspicuous
stimulus
in
their

environment
(18).
Because
the
imprinted
stimulus
is
seen
as
a
parent
or
social
partner,

chickens
prefer
to
spend
time
with
their
imprinted
stimulus
rather
than
a
novel
stimulus.

During
testing,
we
simultaneously
presented
the
newborn
chickens
with
the
Virtual
Parent

on
one
of
the
virtual
walls
and
a
distractor
face
on
the
other
virtual
wall
and
measured
how

much
time
subjects
spent
with
each
stimulus.
This
design
exploits
the
subjects’
natural

motivation
to
spend
time
with
their
parent.
For
each
condition,
the
distractor
face
was

created
by
modifying
the
Virtual
Parent
(See
Figure
1C)
along
a
single
dimension.

Experiment
1
targeted
four
questions
related
to
the
specificity
(3
questions)
and

sensitivity
(1
question)
of
the
first
face
representation
created
by
the
newborn
brain.

1)
Is
the
first
face
representation
selective
for
texture
information
(specificity)?

Conditions
1
&
2
tested
whether
the
Virtual
Parent
could
be
distinguished
from
a
version
of

the
Virtual
Parent
either
with
texture
removed
or
with
only
the
edges
of
the
face
preserved.

Past
studies
with
adult
participants
have
indicated
that
shape
and
texture
information
are

critical
to
face
processing
(21-‐23).
In
these
conditions,
the
subjects
needed
to
rely
on

changes
in
texture
information
alone
to
identify
the
Virtual
Parent.

2)
Is
the
first
face
representation
subject
to
configural
effects
(specificity)?

Conditions
3
&
4
tested
how
feature
information
is
used
in
object
recognition.
Face

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

8

perception
in
adults
is
subject
to
a
“configural
effect,”
which
relies
on
features
remaining
in

the
same
location
and
context
(24).
In
Condition
3,
the
distractor
face
consisted
of
features

alone,
removing
all
feature
context.
Conversely,
in
Condition
4,
the
distractor
face
was

created
by
altering
the
location
of
the
facial
features,
while
holding
the
facial
context

constant.
Configural
effects
have
also
been
implicated
in
findings
that
inversion

significantly
impairs
face
recognition
(25,
26).
In
Condition
5,
subjects
needed
to

distinguish
between
the
Virtual
Parent
and
the
same
face
inverted.

3)
Does
the
first
face
representation
recognize
identity-‐transforming
changes
to

the
face
(specificity)?
The
aim
of
face
processing
is
to
recognize
the
same
face
across

changes
in
the
retinal
image,
while
distinguishing
faces
of
different
individuals.
We
tested

subjects
using
distractor
faces
that
altered
the
Virtual
Parent
along
an
identity-‐
transforming
dimension.
Conditions
6-‐8
tested
versions
of
the
Virtual
Parent
that
were

elderly,
masculine
in
coloring,
or
masculine
in
shape,
respectively.

4)
Is
the
first
face
representation
tolerant
to
identity-‐preserving
changes
to
the

face
(sensitive)?
In
Conditions
9
and
10,
we
tested
whether
subjects
would
discriminate

the
distractor
face
when
the
Virtual
Parent
was
changed
along
an
identity-‐preserving

dimension:
emotional
expression.
In
these
conditions,
the
distractor
faces
were
identical
to

the
Virtual
Parent
but
either
angry
or
fearful,
altering
the
retinal
image
produced,
but
not

the
actual
identity
of
the
face.

Our
second
experiment
tested
whether
the
first
face
representation
created
by
a

newborn
visual
system
can
be
generative,
allowing
for
recognition
in
novel
viewing

conditions.
In
Experiment
2,
a
new
set
of
subjects
were
presented
with
the
same
Virtual

Parent
used
in
Experiment
1;
however,
the
Virtual
Parent
was
shown
from
a
limited

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

9

viewing
range,
rotating
through
an
angle
of
±10º
about
a
vertical
axis
passing
through
its

centroid
(rather
than
±90º
as
in
Experiment
1).
During
the
two
Test
Phases,
subjects

needed
to
distinguish
between
the
Virtual
Parent
and
a
distractor.
Test
Phase
1
used
the
No

Texture
distractor,
while
Test
Phase
2
used
the
Masculine
Color
distractor.
Unlike

Experiment
1,
the
retinal
images
produced
by
the
Virtual
Parent
during
the
Test
Phases
did

not
match
the
retinal
images
produced
during
the
Input
Phase.
During
the
test
trials,
the

Virtual
Parent
was
shown
rotating
10°
through
novel
viewpoints
vertically,
horizontally,
or

diagonally
(rather
than
rotating
horizontally
around
a
vertical
axis,
as
in
the
Input
Phase).

Thus,
to
succeed,
subjects
needed
to
build
a
view-‐invariant
face
representation.

Subjects’
movements
were
tracked
by
micro-‐cameras
embedded
in
the
ceilings
of

the
chambers
and
analyzed
with
automated
animal
tracking
software.
Test
trials
were

scored
as
“correct”
when
subjects
spent
a
greater
proportion
of
time
with
their
imprinted

object
and
“incorrect”
when
they
spent
a
greater
proportion
of
time
with
the
unfamiliar

object.

Results

We
used
hierarchical
Bayesian
models
to
assess
subjects’
performance.
This
analysis

takes
into
account
the
hierarchical
dependencies
inherent
to
the
data
and
the
inter-‐
and

intra-‐subject
variability
for
each
condition.
Here,
we
report
the
actual
probability
that

performance
in
each
condition
is
at
chance
or
lower,
rather
than
traditional
p-‐values
(more

details
on
the
analysis
can
be
found
in
SI).
All
analyses
were
performed
using
R
version

2.15.0
(http://www.r-‐project.org/),
JAGS
(http://mcmc-‐jags.sourceforge.net/)
and

adaptation
of
program
code
from
Dr.
John
Kruschke
(27).
The
subjects
were
able
to

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

10

distinguish
the
Virtual
Parent
from
distractors
despite
having
no
visual
experience
with

any
faces
or
objects
other
than
the
Virtual
Parent
prior
to
testing.

Experiment
1

See
Figure
2
for
results
across
all
trials.

Figure
2

Percent
of
successful
trials
for
each
condition.
The
percent
of
successful
trials
was

calculated
by
collapsing
trials
across
all
subjects
(using
the
raw
data
without
Bayesian

analysis).
Based
on
the
hierarchical
Bayesian
analysis
of
performance,
subjects
were
able
to

distinguish
their
Virtual
Parent
from
the
distractor
in
all
conditions
except
changes
in

emotional
expressions
and
changes
in
feature
configuration.

Subjects
were
successful
at
distinguishing
their
Virtual
Parent
from
a
distractor
face

in
which
all
texture
information
was
removed.
The
probability
that
the
underlying
group

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

11

performance
was
at
chance
(50%)
or
below
was
<0.001
for
both
Conditions
1
and
2.
Thus,

subjects
were
able
to
distinguish
the
Virtual
Parent
from
distractor
faces
with
an
identical

structure.

Feature
context
was
important
for
subjects’
face
recognition.
In
Condition
3

(Features
Alone),
the
probability
of
underlying
group
performance
at
chance
or
below
was

<0.001.
However,
subjects
did
not
successfully
use
featural
spacing
to
recognize
their

Virtual
Parent.
In
Condition
4
(New
Feature
Locations),
the
probability
that
the
underlying

group
performance
was
at
chance
or
below
was
0.919.
These
findings
suggest
that
the

building
blocks
of
the
configural
effect
in
the
newborn
brain
may
not
rely
on
featural

spacing.

Conversely,
the
inversion
effect
was
important
in
face
recognition.
In
Condition
5

(Inverted
Face),
the
probability
that
the
underlying
group
performance
was
at
chance
or

lower
was
0.003.
This
finding
is
especially
striking
given
that
the
frames
of
each
video
(the

Virtual
Parent
and
the
distractor)
were
identical
except
for
their
orientation
in
the
picture

plane.

Subjects
were
also
successful
at
distinguishing
their
Virtual
Parent
from
distractor

faces
with
new
identities.
The
probability
that
the
underlying
group
performance
was
at

chance
or
lower
in
Condition
6
(Elderly
Face)
was
0.001.
The
probability
that
the

underlying
group
performance
was
at
chance
or
lower
in
both
Conditions
7
and
8

(Masculine
Color
and
Masculine
Shape)
was
0.028.

Finally,
subjects
did
not
distinguish
the
Virtual
Parent
from
a
distractor
with
a

different
emotional
expression.
The
probability
that
the
underlying
group
performance

was
at
chance
or
lower
in
Condition
9
(angry
expression)
was
0.431.
The
probability
that

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

12

the
underlying
group
performance
was
at
chance
or
lower
in
Condition
10
(fearful

expression)
was
0.425.
Thus,
just
like
face
processing
in
human
adults,
newborn
chickens

can
recognize
faces
across
identity-‐preserving
transformations
caused
by
facial

expressions.

Experiment
2

When
paired
with
either
distractor,
subjects
were
able
to
recognize
the
Virtual

Parent
across
substantial
changes
in
head
orientation.
The
results
are
shown
in
Figure
3.

Six
hierarchical
Bayesian
analyses
were
performed,
with
conditions
grouped
by
the

distractor
face
(No
Texture
or
Masculine
Color)
and
by
the
extremity
of
the
Virtual
Parent’s

viewpoint
rotation
(0°,
25°,
or
50°).
For
the
No
Texture
distractor,
the
probability
that

performance
was
at
or
below
chance
was
<
0.001
for
all
viewing
angles.

For
the
Masculine

Color
distractor,
the
probability
that
performance
was
at
or
below
chance
was
0.001,

<0.001,
and
0.001
for
0°,
25°,
or
50°,
respectively.
Thus,
the
face
representations
were

remarkably
generative,
i.e.,
insensitive
to
large
(previously
unobserved)
changes
in
head

orientation.

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

13

Figure
3

Heatmaps
depicting
percent
of
successful
trials
for
each
condition.
The
percent
of

successful
trials
was
calculated
by
collapsing
trials
across
all
subjects
(using
the
raw
data

without
Bayesian
analysis).

Discussion

This
study
examined
the
developmental
origins
of
face
recognition
by
probing
the

nature
of
the
first
face
representation
created
by
a
newborn
visual
system.
We
found

evidence
that
the
first
face
representation
created
by
the
newborn
brain
can
be
highly

specific,
sensitive,
and
generative.
Subjects
were
able
to
discriminate
their
Virtual
Parent

from
faces
that
had
undergone
identity-‐transforming
changes,
but
could
not
discriminate

their
Virtual
Parent
from
faces
that
had
undergone
identity-‐preserving
changes.
Moreover,

!"#$%&'$(%)*&'+,)
!"#$%&'()*+),-%.)*
!"#$%&'$(%)*&'+,)
!"#$%&'()*+),-%.)*
/0*
1//0*
2/0*
FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

14

this
first
face
representation
contained
elements
found
in
mature
face
processing,
such
as

reliance
on
texture
and
structure
information,
facial
context,
and
upright
presentation
in

the
picture
plane.
Finally,
our
results
indicate
that
face
representations
are
highly

generative,
allowing
for
recognition
of
the
same
face
across
novel
rotations.

Previous
work
has
studied
the
foundations
of
face
recognition
in
primates
with
no

prior
visual
experience
of
faces
(13).
These
primates
were
able
to
discriminate
faces
that

differed
in
features
and
featural
spacing.
However,
these
subjects’
face
recognition
abilities

may
have
been
heavily
shaped
by
their
extensive
experiences
with
other
visual
objects.
In

our
experiment,
subjects
were
only
exposed
to
the
Virtual
Parent
prior
to
testing,
without

any
exposure
to
other
objects
or
faces.
This
method
allows
for
an
exploration
of
the

experience-‐independent
building
blocks
of
face
recognition.

We
found
evidence
for
some
of
the
signatures
of
adult
face
recognition
in
our

newborn
subjects.
Shape
and
texture
information
are
important
cues
for
face
processing
in

adults
(21-‐23).
The
newborn
chickens
were
able
to
discriminate
the
Virtual
Parent
from

distractors
that
had
a
different
texture,
but
identical
structure
(the
conditions
with
no

texture
information,
only
edges
preserved,
and
masculine
coloring).
Similarly,
the
subjects

were
able
to
discriminate
the
Virtual
Parent
from
a
distractor
with
identical
texture,
but

altered
structure
(the
masculine
shaped
face).
Thus,
sensitivity
to
texture
and
structural

patterns
of
faces
may
be
inherent
to
face
recognition.

Another
property
of
mature
face
recognition
is
configural,
or
holistic
processing.
In

the
current
study,
subjects
showed
some
of
the
hallmarks
of
configural
effects
(Diamond
&

Carey,
1986)
(24,
28).
Subjects
were
sensitive
to
whether
the
face
was
presented
upright
or

inverted.
They
were
also
sensitive
to
whether
the
features
were
presented
in
the
context
of

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

15

a
face
background.
This
suggests
that
the
inversion
effect
and
the
importance
of
feature

context
may
be
part
of
the
building
blocks
for
the
configural
effect
seen
in
mature
face

processing.
However,
subjects
were
not
able
to
discriminate
the
Virtual
Parent
from
a

distractor
that
had
features
warped
to
new
locations.
One
possible
interpretation
of
this

failure
is
that
the
importance
of
featural
spacing
requires
maturation
and
environmental

learning.
This
interpretation
is
consistent
with
studies
finding
that
the
configural
effect

takes
longer
to
develop
than
feature
recognition
(29).
However,
the
finding
is
also

consistent
with
recent
research
claiming
that
face
perception
is
actually
piecemeal:

recognition
of
faces
does
not
outperform
a
Bayesian
algorithm
that
integrates
information

from
each
individual
feature
(30).
This
theory
is
consistent
with
our
finding
that
the

newborn
chickens
did
not
discriminate
a
face
with
the
same
features
moved
to
new

locations.

Finally,
are
newborn
brains
equipped
to
distinguish
identity-‐preserving
changes

from
identity-‐transforming
changes?
A
single
face
can
produce
an
infinite
number
of
retinal

images
(depending
on
changes
in
viewpoint,
position,
lighting,
etc.).
Nonetheless,
the

subjects
were
able
to
discriminate
the
Virtual
Parent
from
distractors
that
varied
by
age

and
gender
(identity-‐transforming
manipulations),
but
not
from
distractors
that
varied
by

emotional
expression
(identity-‐preserving
manipulations).
This
finding
indicates
that
even

the
first
face
representation
created
in
the
newborn
brain
is
subject
to
a
tradeoff
between

sensitivity
(being
tolerant
enough
to
allow
flexibility
to
changes
in
the
retinal
input)
and

specificity
(the
ability
to
recognize
a
change
in
face
identity)
(31).

Importantly,
subjects
were
also
able
to
recognize
the
Virtual
Parent
despite
changes

in
the
orientation
of
the
face.
Previous
theories
of
object
recognition
have
assumed
3-‐D

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

16

object
recognition
is
accomplished
by
associating
together
multiple
views
of
an
object

through
visual
experience
(32).
However,
more
recent
evidence
indicates
that
mature

subjects
are
able
to
recognize
objects
at
novel,
previously
unseen
viewpoints,
up
to
60°
in

rotation
(33).
Similarly,
our
findings
also
indicate
that
face
representations
are
generative

without
requiring
actual
visual
experience
of
a
specific
viewpoint
for
recognition.

Materials
and
Methods

Subjects

Subjects
were
domestic,
newborn
chickens
(Gallus
gallus)
hatched
in
complete

darkness.
Fertilized
eggs
were
obtained
from
a
local
egg
distributor.
Eggs
were
placed
in
an

incubator
in
the
laboratory
until
day
19
of
incubation.
Temperature
was
maintained
at

99.6°F
and
humidity
was
maintained
at
45%.
On
day
19,
we
increased
the
humidity
to
60%.

On
day
1
of
life,
chickens
were
taken
from
the
dark
incubator
room
and
moved
in
complete

darkness
to
their
virtual
reality
chambers
(with
experimenters
using
night
vision
goggles

for
aid).

Materials

The
chickens
were
raised
in
virtual
reality
chambers
(66cm
length,
42cm
width,

69cm
height)
with
two
virtual
walls
(40.5cm
length,
26cm
height)
opposite
each
other.

Food
and
water
were
provided
ad
libitum
in
food
compartments
(2.5cm
width,
66cm

length,
2.7cm
height)
within
the
chambers.
The
food
compartments
were
below
the

subjects
so
that
subjects
were
eating
and
drinking
from
holes
in
the
ground
rather
than

from
object-‐like
containers.
Food
was
a
sand-‐like
grain
because
grain
does
not
behave
with

the
property
of
objects
(i.e.,
solidity
and
rigid
boundaries).
Each
food
compartment
was

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

17

divided
into
thirds
(21cm,
24cm,
and
21cm
in
length)
with
water
in
the
inner

compartments
and
food
in
the
outer
compartments.
Each
chamber
had
flooring
units

through-‐out
(39.5cm
width,
66cm
length,
2.7cm
height)
for
subjects
to
walk
on
that

allowed
waste
to
fall
through,
below
the
subjects.
Each
subjects’
entire
experience
in
the

virtual
reality
chambers
was
tracked
and
monitored
using
Ethovision
XT
7.0
(Noldus

Information
Technology™
Wageningen
,
The
Netherlands)
and
cameras
with
diameter
of

1.5cm.

Imprinting
stimuli
consisted
of
computer-‐presented
animation
of
a
Caucasian

female
face
(ear-‐to-‐ear
width:
6.5cm,
height:
10cm,
distance
about
flooring
unit:
1cm,
see

Fig.
1C).
Imprinting
and
test
animations
were
created
using
frames
from
FaceGen.
At
the

end
of
the
each
trial,
a
black
screen
appeared
for
1
minute
before
the
next
trial
began.

Bayesian
Analysis

Due
to
the
hierarchical
dependencies
in
the
data
collected
(see
SI),
the
data
were

analyzed
using
hierarchical
Bayesian
models.
A
major
advantage
of
using
a
Bayesian

analysis
instead
of
traditional
null
hypothesis
testing
statistics
is
that
Bayesian
models
are

designed
to
be
appropriate
to
the
data
structure
(27,
34),
allowing
for
richer
inferences

from
the
data.
The
model
used
here
provides
parameter
estimates
for
each
individual

newborn
chicken
as
well
as
the
underlying
probability
of
success
for
each
condition

(provided
in
SI).
With
the
Bayesian
analysis,
we
can
calculate
the
actual
probability
that
the

likelihood
of
success
is
50%
or
less
(a
highly
intuitive
statistic
to
interpret),
rather
than
a
p-‐
value
that
calculates
the
probability
of
getting
data
as
extreme
as
the
data
actually
obtained

assuming
that
the
null
hypothesis
is
actually
true.

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

18

For
each
condition,
the
Bayesian
model
uses
an
“uninformative
prior”
of
1
success

and
1
failure
for
the
overall
group
mean.
This
prior
assumes
that
all
means
are
equally

likely,
a
conservative
assumption.
The
model
also
includes
a
parameter
called
kappa
that

represents
“certainty”
(conceptually,
how
consistent
results
are
across
subjects).
We
use
a

uniform
prior
(35)
from
0.000001
to
the
maximum
reasonable
kappa.
The
maximum

reasonable
kappa
is
the
kappa
for
subject
performance
when
differentiating
between
the

Virtual
Parent
and
a
blank
screen,
calculated
during
non-‐test
trials
during
the
Test
Phase.

The
hierarchical
Bayesian
model
uses
Markov
Chain
Monte
Carlo
(MCMC)
sampling

to
approximate
the
posterior
distribution
of
the
parameters
for
each
individual
subject
and

the
hyperparameter
for
each
condition.
The
analysis
used
a
burn-‐in
of
10,000
steps,
with
a

total
of
100,000
steps
after
burn-‐in.

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

19

References

1.
Johnson
MH,
Dziurawiec
S,
Ellis
H,
&
Morton
J
(1991)
Newborns
Preferential

Tracking
of
Face-‐Like
Stimuli
and
Its
Subsequent
Decline.
Cognition
40(1-‐2):1-‐19.

2.
Rosa-‐Salva
O,
Regolin
L,
&
Vallortigara
G
(2010)
Faces
are
special
for
newly
hatched

chicks:
evidence
for
inborn
domain-‐specific
mechanisms
underlying
spontaneous

preferences
for
face-‐like
stimuli.
Developmental
Sci
13(4):565-‐577.

3.
Balas
B
(2010)
Using
innate
visual
biases
to
guide
face
learning
in
natural
scenes:
a

computational
investigation.
Developmental
Sci
13(3):469-‐478.

4.
Johnson
MH
(2005)
Subcortical
face
processing.
Nat
Rev
Neurosci
6(10):766-‐774.

5.
Kanwisher
N
(2000)
Domain
specificity
in
face
perception.
Nature
neuroscience

3(8):759-‐763.

6.
Kanwisher
N
&
Yovel
G
(2006)
The
fusiform
face
area:
a
cortical
region
specialized

for
the
perception
of
faces.
Philos
T
R
Soc
B
361(1476):2109-‐2128.

7.
Gauthier
I,
et
al.
(2000)
The
fusiform
"face
area"
is
part
of
a
network
that
processes

faces
at
the
individual
level.
J
Cognitive
Neurosci
12(3):495-‐504.

8.
Pitcher
D,
Walsh
V,
Yovel
G,
&
Duchaine
B
(2007)
TMS
evidence
for
the
involvement

of
the
right
occipital
face
area
in
early
face
processing.
Curr
Biol
17(18):1568-‐1573.

9.
Geldart
S,
Mondloch
CJ,
Maurer
D,
de
Schonen
S,
&
Brent
HP
(2002)
The
effect
of

early
visual
deprivation
on
the
development
of
face
processing.
Developmental
Sci

5(4):490-‐501.

10.
Le
Grand
R,
Mondloch
CJ,
Maurer
D,
&
Brent
HP
(2003)
Expert
face
processing

requires
visual
input
to
the
right
hemisphere
during
infancy.
Nature
neuroscience

6(10):1108-‐1112.

11.
Le
Grand
R,
Mondloch
CJ,
Maurer
D,
&
Brent
HP
(2004)
Impairment
in
holistic
face

processing
following
early
visual
deprivation.
Psychological
science
15(11):762-‐768.

12.
Pascalis
O,
Deschonen
S,
Morton
J,
Deruelle
C,
&
Fabregrenet
M
(1995)
Mothers
Face

Recognition
by
Neonates
-‐
a
Replication
and
an
Extension.
Infant
Behav
Dev

18(1):79-‐85.

13.
Sugita
Y
(2008)
Face
perception
in
monkeys
reared
with
no
exposure
to
faces.

Proceedings
of
the
National
Academy
of
Sciences
of
the
United
States
of
America

105(1):394-‐398.

14.
Edelman
S
&
Intrator
N
(2003)
Towards
structural
systematicity
in
distributed,

statically
bound
visual
representations.
Cognitive
Sci
27(1):73-‐109.

15.
Olshausen
BA
&
Field
DJ
(1996)
Emergence
of
simple-‐cell
receptive
field
properties

by
learning
a
sparse
code
for
natural
images.
Nature
381(6583):607-‐609.

16.
Wallis
G
&
Rolls
ET
(1997)
Invariant
face
and
object
recognition
in
the
visual

system.
Prog
Neurobiol
51(2):167-‐194.

17.
Oreilly
RC
&
Johnson
MH
(1994)
Object
Recognition
and
Sensitive
Periods
-‐
a

Computational
Analysis
of
Visual
Imprinting.
Neural
Comput
6(3):357-‐389.

18.
Bolhuis
JJ
(1999)
Early
learning
and
the
development
of
filial
preferences
in
the

chick.
Behavioural
brain
research
98(2):245-‐252.

19.
Fontanari
L,
Rugani
R,
Regolin
L,
&
Vallortigara
G
(2011)
Object
individuation
in
3-‐
day-‐old
chicks:
use
of
property
and
spatiotemporal
information.
Developmental
Sci

14(5):1235-‐1244.

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

20

20.
Wang
YA,
Brzozowska-‐Prechtl
A,
&
Karten
HJ
(2010)
Laminar
and
columnar

auditory
cortex
in
avian
brain.
Proceedings
of
the
National
Academy
of
Sciences
of
the

United
States
of
America
107(28):12676-‐12681.

21.
Hill
H,
Bruce
V,
&
Akamatsu
S
(1995)
Perceiving
the
Sex
and
Race
of
Faces
-‐
the
Role

of
Shape
and
Color.
P
Roy
Soc
Lond
B
Bio
261(1362):367-‐373.

22.
Said
CP
&
Todorov
A
(2011)
A
Statistical
Model
of
Facial
Attractiveness.

Psychological
science
22(9):1183-‐1190.

23.
Walker
M
&
Vetter
T
(2009)
Portraits
made
to
measure:
Manipulating
social

judgments
about
individuals
with
a
statistical
face
model.
J
Vision
9(11).

24.
Farah
MJ,
Wilson
KD,
Drain
M,
&
Tanaka
JN
(1998)
What
is
"special"
about
face

perception?
Psychol
Rev
105(3):482-‐498.

25.
Diamond
R
&
Carey
S
(1986)
Why
Faces
Are
and
Are
Not
Special
-‐
an
Effect
of

Expertise.
J
Exp
Psychol
Gen
115(2):107-‐117.

26.
Yin
RK
(1969)
Looking
at
Upside-‐down
Faces.
J
Exp
Psychol
81(1):141-‐&.

27.
Kruschke
JK
(2010)
What
to
believe:
Bayesian
methods
for
data
analysis.
Trends
in

cognitive
sciences
14(7):293-‐300.

28.
Maurer
D,
Le
Grand
R,
&
Mondloch
CJ
(2002)
The
many
faces
of
configural

processing.
Trends
in
cognitive
sciences
6(6):255-‐260.

29.
Mondloch
CJ,
Le
Grand
R,
&
Maurer
D
(2002)
Configural
face
processing
develops

more
slowly
than
featural
face
processing.
Perception
31(5):553-‐566.

30.
Gold
JM,
Mundy
PJ,
&
Tjan
BS
(2012)
The
Perception
of
a
Face
Is
No
More
Than
the

Sum
of
Its
Parts.
Psychological
science
23(4):427-‐434.

31.
Serre
T,
Oliva
A,
&
Poggio
T
(2007)
A
feedforward
architecture
accounts
for
rapid

categorization.
Proceedings
of
the
National
Academy
of
Sciences
of
the
United
States

of
America
104(15):6424-‐6429.

32.
Stryker
MP
(1991)
Neurobiology
-‐
Temporal
Associations.
Nature
354(6349):108-‐
109.

33.
Wang
G,
Obama
S,
Yamashita
W,
Sugihara
T,
&
Tanaka
K
(2005)
Prior
experience
of

rotation
is
not
required
for
recognizing
objects
seen
from
different
angles.
Nature

neuroscience
8(12):1768-‐1775.

34.
Kruschke
JK
(2011)
Doing
bayesian
data
analysis
:
a
tutorial
with
R
and
BUGS

(Academic
Press,
Burlington,
MA)
pp
xvii,
653
p.

35.
Gelman
A
(2006)
Prior
distributions
for
variance
parameters
in
hierarchical

models(Comment
on
an
Article
by
Browne
and
Draper).
Bayesian
Anal
1(3):515-‐
533.

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

21

Supplementary
Information
Contents
&
Captions

• Figure
1:
Schematic
of
hierarchical
Bayesian
model.
Top
diagram
is
a
Bayes
Net

diagram
depicting
the
hierarchical
dependency
of
the
data.
Bottom
diagram

illustrates
how
the
hierarchical
dependencies
are
incorporated
in
the
hierarchical

Bayesian
analysis.
The
analysis
estimates
the
hyperparameter
of
each
condition
and

the
parameter
for
each
individual
chick.

• Figures
2
-‐
11:
Experiment
1
Results
(overall
and
by
subject)
of
hierarchical

Bayesian
analysis.
Each
figure
shows
the
distractor
face
alongside
the
Virtual

Parent.
The
data
are
shown
in
probability
density
charts
for
the
hyperparameter
of

the
condition
and
the
parameter
for
each
individual
chick.

• Figure
12:
Experiment
1
time
course
of
test
phase
results
graphed
as
performance

by
day.

• Figures
13
-‐
18:
Experiment
2
Results
(overall
and
by
subject)
of
hierarchical

Bayesian
analysis.
Each
figure
shows
the
distractor
face
alongside
the
different
test

head
orientations
of
the
Virtual
Parent.
The
data
are
shown
in
probability
density

charts
for
the
hyperparameter
of
the
condition
and
the
parameter
for
each

individual
chick.

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

22

SI
Figure
1

!"#$%&
'&
!"#$%&
(&
)*+,-./&
'&
0123#412&
#&
!"#$%&
5&
!"#$%&
'&
!"#$%&
(&
)*+,-./&
(&
!"#$%&
5&
!"#$%&
'&
!"#$%&
(&
)*+,-./&
,&
!"#$%&
5&
!"#"$%&'("()*&)"+$!','-.'-)&'/$
!"#$%&"#' 6'7&'8&
92#:1";&<#=/"#+*412&
6>?>>>>>'7&&
-=/?&;$@#;*;8&
AB& 6'CA8B&
6$7&+8&
()*+'!+%*,'
D-"21*%%#&
<#=/"#+*412&
012304$!020$
D-/$&<#=/"#+*412&
D-/$&<#=/"#+*412&
-%./)/*+%*)0'1)2.3%)#'4"$.0'
FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

23

SI
Figure
2

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

24

SI
Figure
3

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

25

SI
Figure
4

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

26

SI
Figure
5

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

27

SI
Figure
6

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

28

SI
Figure
7

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

29

SI
Figure
8

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

30

SI
Figure
9

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

31

SI
Figure
10

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

32

SI
Figure
11

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

33

SI
Figure
12

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

34

SI
Figure
13

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

35

SI
Figure
14

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

36

SI
Figure
15

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

37

SI
Figure
16

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

38

SI
Figure
17

FACE
RECOGNITION
IN
A
NEWBORN
VISUAL
SYSTEM

39

SI
Figure
18

Asset Metadata

Creator Waters, Samantha M. (author)

Core Title Sensitive, specific, and generative face recognition in a newborn visual system

Contributor Electronically uploaded by the author (provenance)

School College of Letters, Arts and Sciences

Degree Master of Arts

Degree Program Psychology

Publication Date 04/18/2013

Defense Date 10/08/2012

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag chicken,cognition,Development,early experience,face recognition,OAI-PMH Harvest

Format application/pdf (imt)

Language English

Advisor Bechara, Antoine (committee member), Mintz, Toben H. (committee member), Tjan, Bosco S. (committee member)

Creator Email samantha.waters@gmail.com,samantha.waters@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-239134

Unique identifier UC11294791

Identifier etd-WatersSama-1548.pdf (filename),usctheses-c3-239134 (legacy record id)

Legacy Identifier etd-WatersSama-1548.pdf

Dmrecord 239134

Document Type Thesis

Format application/pdf (imt)

Rights Waters, Samantha M.

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Abstract (if available)

Abstract Despite the computational difficulty, adults easily recognize faces with incredibly high sensitivity and specificity across novel changes in viewing conditions. Mature face representations are invariant to identity-preserving changes (e.g., changes in expression or viewing angle) while intolerant to identity-transforming changes (e.g., changes in texture, age, and gender). Yet, due to constraints in testing newborn subjects, the origins of these characteristics remain unknown. Are highly sensitive, specific, and generative face representations the product of extensive visual experience, or do genes build neural machinery that generates sensitive, specific, and generative face representations from the onset of experience with faces? To probe the nature of the first face representation created in the newborn mind, we reared newborn chickens for one week in a virtual reality chamber that contained no objects except a single virtual human face (the “Virtual Parent”). Immediately after, the chickens were tested for one week on their ability to discriminate between the Virtual Parent and a set of distractor faces. Distractors were identical to the Virtual Parent, with the exception of a single dimension (e.g., changes in texture, structure, inversion, and expressions). Subjects were able to distinguish the Virtual Parent from distractors that differed in texture, age, gender, and inversion, but not distractors that differed in feature configuration or facial expression. Further, subjects were able to recognize the Virtual Parent across novel changes in head orientation (i.e., view direction), indicating that subjects built invariant face representations. These results show that the first face representation built by the newborn visual system can be sensitive, specific, and generative.