The development of object recognition in the newborn brain

PDF

Download

Open document

Transcript (if available)

Content

THE
DEVELOPMENT
OF
OBJECT
RECOGNITION
IN
THE
NEWBORN
BRAIN

by

Samantha
M.
W.
Wood

A
Dissertation
Presented
to
the

FACULTY
OF
THE
USC
GRADUATE
SCHOOL

UNIVERSITY
OF
SOUTHERN
CALIFORNIA

In
Partial
Fulfillment
of
the

Requirements
for
the
Degree

DOCTOR
OF
PHILOSOPHY

(PSYCHOLOGY)

August
2017

Copyright
2017

Samantha
M.
W.
Wood

ii

Acknowledgements

I
would
like
to
thank
all
of
my
committee
members,
Antoine
Bechara,
Toben
Mintz,

Irving
Biederman,
Stephen
Read,
and
Laurent
Itti,
for
their
considerable
insights
and

guidance.
In
particular,
Antoine
has
provided
valuable
mentorship
during
my
graduate

studies.
I
have
learned
so
much
from
Antoine
about
the
brain,
decision-‐making,
and
the

general
field
of
academics.
I
am
extremely
grateful
for
his
time
and
support.
I
would
also

like
to
thank
my
collaborators
Susan
Schembre,
Qinghua
He,
Lin
Xiao,
Jeffrey
Engelmann,

Jason
Goldman,
and
Aditya
Prasad,
as
well
as
research
assistants
and
lab
managers
who

have
contributed
to
my
research
efforts:
Alex
Hollihan,
Stephanie
Castillo,
and
Lynette
Tan.

Thank
you
to
the
friends
I
have
made
in
graduate
school,
especially
Vanessa
Singh,
who
has

proven
to
be
a
friend
across
careers
and
time
zones.

I
also
offer
sincere
gratitude
to
Justin
Wood
as
my
collaborator,
partner,
and
best

friend.
Justin’s
infectious
curiosity
about
the
mind
is
inspiring
to
experience.
Throughout

my
time
in
graduate
school,
Justin
has
reminded
me
of
my
capabilities
when
I
needed

encouragement,
and
held
me
to
my
capabilities
as
a
mentor.
I
thank
Justin
for
his
unending

support
and
for
helping
me
grow
as
a
scholar.

Finally,
I’d
like
to
dedicate
this
work
to
the
two
most
important
women
in
my
life.

First,
to
my
mother,
Joan
McCartan,
who
instilled
in
me
a
love
of
learning.
My
mother

taught
me
that
math
problems
are
puzzles
to
unravel,
that
books
can
transport
us
to
new

worlds
and
times,
and
that
“smart
kids
are
never
bored.”
From
my
mother,
I
learned
that

“the
good
life
is
one
inspired
by
love
and
guided
by
knowledge”
(Bertrand
Russell,
What
I

Believe).
Second,
to
my
daughter,
Mackenzie
Waters
Wood,
who
I
hope
will
be
inspired
by

iii

my
studies
in
psychology
to
value
education
and
to
always
try
to
understand
others.

Mackenzie,
the
world
is
full
of
wonders
and
riddles
waiting
to
be
explored.

iv

TABLE
OF
CONTENTS

ABSTRACT

1

CHAPTER
1:
Introduction

2

CHAPTER
2:

Newborn
chicks
segment
objects
from
backgrounds
at
the
onset

of
vision

20

Abstract

20

Introduction

21

Methods

24

Results

30

Discussion

34

CHAPTER
3:
The
development
of
background-‐invariant
object
recognition
in

visually
naïve
animals

37

Abstract

37

Introduction

38

Methods

41

Results

45

Discussion

51

CHAPTER
4:
Newborn
chicks
generate
view-‐invariant
object
representations

from
sparse
visual
input

53

Abstract

53

Introduction

54

Methods

57

Results

61

Discussion

75

v

CHAPTER
5:
Face
recognition
in
newborn
chicks
at
the
onset
of
vision

80

Abstract

80

Introduction

81

Methods

85

Results

89

Discussion

94

CHAPTER
6:
A
slowness
constraint
on
the
development
of
view-‐invariant

face
recognition

99

Abstract

99

Introduction

100

Experiment
1

103

Experiment
2

110

Experiment
3

112

Experiment
4

119

General
Discussion

121

CHAPTER
7:
Conclusion

124

REFERENCES

133

1

Abstract

A
central
goal
in
psychology
and
neuroscience
is
to
understand
how
biological
visual

systems
recognize
objects.
However,
the
developmental
origins
of
object
recognition

remain
poorly
understood.
What
object
recognition
abilities
are
present
at
the
onset
of

vision,
and
what
visual
experiences
are
necessary
to
develop
these
abilities?
To
address

these
questions,
my
dissertation
used
an
automated
controlled-‐rearing
method
with

newborn
chicks.
Chapters
2
and
3
examined
the
development
of
background-‐invariant

object
recognition
in
newborns.
These
studies
showed
that
newborn
chicks
can
begin

building
background-‐invariant
object
representations
at
the
onset
of
vision,
and
that
the

development
of
this
ability
requires
visual
experience
with
objects
moving
on
patterned

backgrounds.
Chapter
4
demonstrated
that
newborn
chicks
can
begin
building
view-‐
invariant
representations
of
objects
at
the
onset
of
vision,
and
that
these
abstract

representations
can
be
built
from
sparse
visual
input
(as
little
as
three
views
of
an
object).

Chapter
5
showed
that
newborn
chicks
are
capable
of
face
recognition
at
the
onset
of

vision.
Finally,
Chapter
6
showed
that
newborn
chicks
can
build
view-‐invariant
face

representations,
and
that
the
development
of
this
ability
requires
visual
experience
with

slowly
moving
faces.
Together
these
studies
show
that
newborns
can
develop
high-‐level

visual
recognition
abilities
rapidly,
within
the
first
few
days
of
life.
However,
these
abilities

do
not
develop
automatically;
rather,
the
development
of
high-‐level
vision
requires
visual

experience
with
a
natural
visual
environment,
containing
objects
and
faces
that
move

slowly
over
time
across
patterned
backgrounds.
These
results
begin
to
reveal
how

foundational
visual
abilities
emerge
in
newborn
brains
as
a
function
of
specific
visual

experiences.

2

Chapter
1:
Introduction

Upon
first
opening
their
eyes,
a
newborn
faces
a
monumental
computational
task:

they
must
transform
streams
of
unstructured
sensory
input
into
meaningful
information

about
the
surrounding
environment.
Although
this
task
feels
effortless
to
human
adults,
the

underlying
mental
computations
are
highly
complex.
Visual
input
to
retinal
cells
is

quantitative
and
continuous,
but
our
percepts
of
objects
are
qualitative
and
abstract.
We

perceive
discrete,
segmented
objects
that
persist
through
visual
transformations.
As
a

result,
understanding
the
initial
state
of
postnatal
vision
and
the
role
of
experience
in

shaping
visual
cognition
is
a
central
question
in
philosophy,
cognitive
science,
and

neuroscience.
While
prior
research
has
revealed
important
insights
into
the
prenatal

development
of
neural
structures
and
the
visual
abilities
of
infants
with
months
of
visual

experience,
little
is
known
about
how
visual
perception
and
cognition
emerge
in
the

newborn
brain.

In
order
to
perceive
objects
successfully,
newborns
must
solve
at
least
two

problems.
First,
the
newborn
visual
system
must
parse
objects
from
the
surrounding
scene.

Visual
input
is
comprised
of
regions
with
different
luminance,
hue,
and
texture
values
that

must
be
segmented
into
meaningful
entities.
Second,
the
visual
system
must
recognize

objects
across
novel
viewing
situations.
This
latter
ability,
known
as
invariant
object

recognition
1
,
underlies
the
perception
that
an
object’s
identity
persists
across
novel

surroundings,
viewpoint
angles,
lighting
conditions,
object
positions,
etc.
These
abilities

1
In this dissertation, I use “invariant” to mean tolerant to substantial image variation. A
representation may be invariant by this definition without being fully invariant (i.e., recognizable
from any novel viewing condition without performance costs).

3

require
abstraction
insofar
as
they
imply
knowledge
beyond
physical
similarities
between

images
of
an
object.
Visual
parsing
and
invariant
recognition
allow
individuals
to
make

inferences
about
the
entities
in
the
environment
that
give
rise
to
the
immediate
visual

input
(Helmholtz
&
Southall,
1924;
Hochberg,
1978;
Spelke,
1990).

Psychologists
have
debated
the
origins
of
object
perception
for
nearly
a
century.

Classical
accounts
of
the
development
of
object
perception
offered
two
competing
theories.

Gestalt
theory
posited
that
the
visual
system
interprets
perceptual
information
by
inferring

the
simplest
and
most
regular
external
environment
that
is
consistent
with
the
given
visual

input
(Koffka,
1935;
Köhler,
1929).
For
example,
objects
in
the
world
tend
to
be
coherent

and
regular,
rather
than
fragmented
and
disorganized.
Because
Gestalt
psychologists

believed
that
these
principles
of
visual
perception
arose
from
an
inherent
tendency
of

neurons
to
expend
minimum
energy
(Koffka,
1935),
they
maintained
that
newborns
must

also
perceive
an
organized
and
coherent
world.
Conversely,
empiricist
theories
argued
that

the
newborn
visual
system
cannot
initially
make
sense
of
visual
input.
Specifically,

empiricists
claimed
that
newborns
cannot
transform
immediate
visual
patterns
into

information
about
the
surrounding
environment
until
they
can
move
around
the
world
and

manipulate
objects
(Berkeley,
1910;
Helmholtz
&
Southall,
1924;
Piaget,
1952).
As
infants

act
upon
objects,
they
learn
how
the
properties
of
visual
input
map
onto
physical

properties
of
objects.
These
competing
theories
have
sparked
heated
debate
in
the

developmental
literature
about
the
origins
of
object
perception.
What
mechanisms
govern

the
perception
of
objects
at
the
onset
of
vision,
and
how
are
those
mechanisms
shaped
by

experience?

4

Summary
of
Previous
Research

Research
with
human
infants

Studies
of
parsing
in
human
infants
have
generally
focused
on
the
boundaries

between
two
objects.
In
one
popular
paradigm,
infants
are
presented
with
two
adjacent

objects
that
have
an
aligned
edge.
A
researcher
then
grasps
and
pulls
on
one
of
the
objects

and
either
both
objects
move
together
as
a
single
unit
(a
“move-‐together
event”)
or
the

pulled
object
moves
while
the
other
object
remains
stationary
(a
“move-‐apart
event”).
If

infants
can
successfully
segment
the
two
objects,
then
they
should
expect
a
move-‐apart

event.
In
another
testing
paradigm,
infants
are
shown
a
partly
occluded
object
moving
back

and
forth.
If
infants
interpret
the
display
as
a
single
coherent
object
moving
behind
an

occluder,
then
they
should
expect
to
see
a
single
continuous
object
when
the
occluder
is

removed,
rather
than
two
separate
objects.
Studies
using
these
methodologies
have
found

that
4-‐month-‐old
infants
are
able
to
parse
the
objects
in
these
displays.
By
4
months
of
age,

infants
can
use
object
features
to
define
object
boundaries
(Kestenbaum,
Termine,
&

Spelke,
1987;
Needham,
2000;
Needham
&
Ormsbee,
2003)
and
perceive
the
unity
of
the

visible
object
parts
(Kellman
&
Spelke,
1983).
Furthermore,
motion
is
particularly

informative
for
determining
object
boundaries
in
this
task
(Kellman,
Spelke,
&
Short,

1986).
These
studies
are
often
interpreted
as
revealing
an
early
emerging
ability
to

understand
boundaries
within
a
scene
and
perceive
segregation
of
figure
and
ground.

However,
to
my
knowledge,
infant
studies
have
not
directly
tested
infants’
ability
to
parse

objects
from
background
scenes
or
visual
clutter.
In
my
dissertation,
I
examine
directly

whether
newborn
animals
are
capable
of
segmenting
objects
from
backgrounds
at
the

onset
of
vision.

5

In
contrast
to
studies
of
visual
parsing,
only
a
few
studies
have
examined
invariant

object
recognition
in
young
human
infants,
and
these
studies
have
produced
mixed
results

(possibly
due
to
differing
task
demands).
An
early
study
reported
that
9-‐month-‐old

infants—but
not
6-‐month-‐old
infants—could
use
object
shape
alone
to
recognize
a
familiar

object
(Ruff,
1978).
Conversely,
Soska
&
Johnson
(2008)
found
evidence
for
three-‐
dimensional
shape
representations
in
6-‐month-‐old
infants.
After
being
habituated
to
an

object
rotating
15°,
the
6-‐month-‐old
infants
perceived
the
object
as
a
solid
(complete)

volume
rather
than
a
hollow
form.
Other
studies
have
reported
that
4-‐month-‐old
infants

can
recognize
objects
from
novel
viewpoints
if
the
original
presentation
of
the
object

provided
kinetic
depth
information
(Kellman,
1984;
Kellman
&
Short,
1987;
Owsley,
1983).

These
studies
suggest
invariant
object
recognition,
like
object
segmentation,
relies
on

motion
information.
In
addition
to
changes
in
viewpoints,
visual
systems
must
also
learn

invariance
to
other
transformations
such
as
changes
in
illumination.
A
recent
study
found

that
3-‐
to
4-‐month-‐old
infants
are
highly
sensitive
to
changes
in
pixel
intensity
caused
by

minute
illumination
changes,
but
by
7-‐
to
8-‐months
of
age
infants
become
less
sensitive
to

illumination
changes
and
more
sensitive
to
changes
in
objects’
surface
properties
(Yang,

Kanazawa,
Yamaguchi,
&
Motoyoshi,
2015)

While
studies
of
human
infants
have
provided
important
insights
about
the

development
of
object
perception
early
in
life,
these
studies
are
subject
to
a
number
of

limitations.
First,
studies
of
human
infants
are
typically
able
to
collect
only
a
small
amount

of
test
data
per
subject.
Thus,
it
has
generally
not
been
possible
to
study
perceptual

development
in
young
infants
with
high
precision.
In
addition,
high
measurement
error

(colloquially,
“noise”)
as
well
as
flexibility
in
stimuli
presentation,
data
coding,
and
subject

6

exclusion
produce
high
false-‐positive
rates
and
reduce
replicability
of
findings
(Loken
&

Gelman,
2017;
Simmons,
Nelson,
&
Simonsohn,
2011).
Second,
human
infants
cannot
be

raised
in
controlled
environments
from
birth.
Even
infants
who
are
just
a
few
months
old

have
already
acquired
hundreds
of
hours
of
patterned
visual
experience
(Johnson,
Amso,
&

Slemmer,
2003).
Thus,
studies
of
human
infants
are
unable
to
examine
(1)
the
initial
state

of
object
perception
(i.e.,
the
state
of
object
recognition
machinery
at
the
onset
of
vision)

and
(2)
how
visual
experience
shapes
that
initial
state
over
time.

Studies
of
patients
recovering
from
blindness

Another
approach
to
understanding
the
development
of
object
perception
has

focused
on
congenitally
blind
individuals
who
have
had
their
sight
restored.
Project

Prakash
(Sinha,
2013)
treats
blind
individuals
in
underprivileged
areas
of
India
and

subsequently
tests
how
these
individuals
make
sense
of
the
new
bombardment
of
visual

input.
In
one
such
study,
newly-‐sighted
patients
(2
weeks
to
3
months
post-‐surgery)
were

tested
on
their
ability
to
parse
objects
that
were
displayed
as
static
images
(Ostrovsky,

Meyers,
Ganesh,
Mathur,
&
Sinha,
2009).
Subjects
were
unable
to
parse
simple,
static

illustrations
of
overlapping
shapes
as
well
as
real-‐world
images
of
objects.
However,
when

the
stimuli
incorporated
motion
cues,
the
patients
were
able
to
parse
the
displays

successfully.
Moreover,
following
a
longer
delay
post-‐treatment
(10-‐18
months
post-‐
surgery),
the
patients
showed
significant
improvement
in
parsing
static
images,
suggesting

that
the
visual
system
can
learn
to
parse
scenes
through
natural
visual
experience.
These

findings
reinforce
the
results
from
studies
of
human
infants
suggesting
that
motion
is
a

critical
cue
for
the
visual
system
to
learn
how
to
parse
objects
in
scenes.

7

The
research
from
Project
Prakash
elegantly
joins
humanitarian
efforts
with

progress
in
basic
science;
however,
there
are
limitations
to
studies
of
congenitally
blind

patients.
First,
like
human
infants,
patients
recovering
from
blindness
acquire
weeks
to

months
of
visual
experiences
prior
to
the
experiments.
Thus,
it
is
not
possible
to
determine

whether
the
visual
abilities
found
during
testing
are
present
at
the
onset
of
vision
or

learned
from
experience
with
a
natural
visual
world.
Second,
visual
deprivation
leads
to

cross-‐modal
reorganization
of
the
visual
cortex
(Collignon
et
al.,
2015;
Maidenbaum,

Abboud,
&
Amedi,
2014).
Therefore,
in
blind
patients,
the
visual
cortex
has
been
shaped
by

the
natural
statistics
of
perceptual
input
from
other
modalities.
The
initial
state
of
vision
in

a
blind
patient
is
not
equivalent
to
the
initial
state
of
vision
in
a
newborn.

Newborn
chicks
as
a
model
system
for
studying
the
development
of
object

perception

Animal
models
provide
a
critical
tool
in
the
investigation
of
visual
processing

machinery.
To
date,
nonhuman
primates
have
been
the
model
of
choice
for
studying
visual

cognition
because
their
visual
systems
closely
mirror
our
own.
Studies
of
primates
have

revealed
many
important
characteristics
about
object
recognition,
including
the
nature
of

its
underlying
computations
and
the
architecture
of
its
neural
substrates
(reviewed
by

DiCarlo,
Zoccolan,
&
Rust,
2012;
see
also
Yamins
et
al.,
2014).
There
is
also
growing

evidence
that
rats
and
pigeons
may
be
promising
animal
models
for
studying
object

recognition
because
they,
too,
have
invariant
object
recognition
abilities
(Alemi-‐Neissi,

Rosselli,
&
Zoccolan,
2013;
Soto,
Siow,
&
Wasserman,
2012;
Tafazoli,
Di
Filippo,
&
Zoccolan,

2012;
Wasserman
&
Biederman,
2012;
Zoccolan,
Oertelt,
DiCarlo,
&
Cox,
2009).
These

8

animal
models
enable
experimental
techniques
that
are
difficult
to
perform
with
primates.

For
instance,
rat
studies
allow
the
application
of
a
wide
range
of
techniques
including

molecular
and
histological
approaches,
two-‐photon
imaging,
and
large-‐scale
recordings

from
multiple
brain
areas.
However,
while
primates,
rodents,
and
pigeons
have
many

advantages
as
model
systems,
these
animals
are
not
well
suited
for
studying
the
initial
state

of
object
recognition
because
they
cannot
be
raised
in
strictly
controlled
environments

from
birth.
2

These
three
animal
models
all
require
parental
care.
Thus,
after
birth
or
hatching,

the
newborns
must
be
raised
in
environments
that
contain
a
caregiver.
Experience
with

this
caregiver
could
significantly
shape
the
newborn’s
object
recognition
mechanisms
by

providing
clues
about
which
retinal
image
changes
are
identity-‐preserving
transformations

and
which
are
not.
Indeed,
studies
of
monkeys
and
humans
show
that
object
recognition

machinery
changes
rapidly
in
response
to
statistical
redundancies
in
the
organism’s

environment
(e.g.,
Cox,
Meier,
Oertelt,
&
DiCarlo,
2005;
Wallis
&
Bülthoff,
2001),
with

significant
neuronal
rewiring
occurring
in
as
little
as
one
hour
of
experience
with
an

altered
visual
world
(Li
&
DiCarlo,
2008,
2010).
There
is
also
extensive
behavioral
evidence

that
primates
begin
encoding
statistical
redundancies
soon
after
birth
(Bulf,
Johnson,
&

Valenza,
2011;
Kirkham,
Slemmer,
&
Johnson,
2002;
Saffran,
Aslin,
&
Newport,
1996).

2
Rats and mice can be reared in darkness. However, dark rearing prevents complete microcircuit
maturation in the visual cortex (Ko, Mrsic-Flogel, & Hofer, 2014), produces abnormalities in
local cortical connectivity (Ishikawa, Komatsu, & Yoshimura, 2014), and alters the long-term
development of GABAergic transmission (Morales, Choi, & Kirkwood, 2002). Further, rats and
mice cannot be raised from birth in controlled, lighted environments (i.e., environments devoid
of objects and agents). In contrast, chicks can be raised in controlled, lighted environments
immediately after hatching. Thus, with chicks, it is possible to examine how patterned visual
input drives the emergence of object recognition at the beginning of the post-embryonic phase of
the animal’s life cycle.

9

These
findings
allow
for
the
possibility
that
even
early
emerging
object
recognition
abilities

(e.g.,
abilities
emerging
days,
weeks,
or
months
after
birth)
are
learned
from
experience

with
objects
early
in
postnatal
life.

Analyzing
the
initial
state
of
visual
cognition
therefore
requires
a
newborn
animal

model
with
two
characteristics:
(1)
the
animal
can
develop
visual
cognitive
abilities
and

(2)
the
animal’s
visual
environment
can
be
strictly
controlled
immediately
after
the
post-‐
embryonic
phase
of
their
life
cycle
(i.e.,
to
prevent
learning
from
visual
experience).

Newborn
chicks
meet
both
of
these
criteria.
First,
chicks
develop
high-‐level
object

recognition
abilities
rapidly
(Wood,
2013,
2015).
For
example,
chicks
can
build
a
view-‐
invariant
representation
of
the
first
object
they
see
in
their
life
(Wood,
2013,
2015).
Chicks

also
have
other
advanced
object
recognition
abilities,
including
the
ability
to
bind
color
and

shape
features
into
integrated
color-‐shape
units
at
the
onset
of
vision
(Wood,
2014).

Second,
chicks
can
be
raised
from
birth
in
environments
devoid
of
objects
and
caregivers

(Vallortigara,
2012;
Wood,
2013).
Unlike
newborn
primates,
rodents,
and
pigeons,

newborn
chicks
do
not
require
parental
care
and
are
immediately
able
to
explore
their

environment.

In
addition,
chicks
imprint
to
objects
seen
soon
after
hatching
(Bateson,
2000;
Horn,

2004).
Chicks
develop
a
strong
attachment
to
their
imprinted
objects,
and
will
attempt
to

spend
most
of
their
time
with
the
objects.
This
imprinting
behavior
can
be
used
to
test

chicks’
object
recognition
abilities
without
supervised
training
(Bolhuis,
1999;
Regolin
&

Vallortigara,
1995;
Wood,
2013).

Notably,
studies
of
chicks
can
also
inform
human
visual
development
because
birds

and
mammals
use
similar
neural
mechanisms.
At
a
macro-‐level,
avian
and
mammalian

10

brains
share
the
same
large-‐scale
organizational
principles:
both
are
modular,
small-‐world

networks
with
a
connective
core
of
hub
nodes
that
includes
prefrontal-‐like
and

hippocampal
structures
(Shanahan,
Bingman,
Shimizu,
Wild,
&
Gunturkun,
2013).
Further,

avian
and
mammalian
brains
have
homologous
cortical-‐like
cells
and
circuits
for

processing
sensory
information
(Dugas-‐Ford,
Rowell,
&
Ragsdale,
2012;
Jarvis
et
al.,
2005;

Karten,
2013;
Wang,
Brzozowska-‐Prechtl,
&
Karten,
2010).
Although
these
neural
circuits

are
organized
differently
in
birds
and
mammals
(nuclear
versus
layered
organization,

respectively),
they
share
many
similarities
in
terms
of
cell
morphology,
the
connectivity

pattern
of
the
input
and
output
neurons,
gene
expression,
and
function
(Butler,
1994;

Karten,
1991,
1997;
Karten
&
Shimizu,
1989;
Medina
&
Reiner,
2000;
Reiner,
Yamamoto,
&

Karten,
2005;
Saini
&
Leppelsack,
1981).
For
instance,
in
chicken
neural
circuitry,
sensory

inputs
are
organized
in
a
radial
columnar
manner,
with
lamina
specific
cell
morphologies,

recurrent
axonal
loops,
and
re-‐entrant
pathways,
typical
of
layers
2–5a
of
mammalian

neocortex
(reviewed
by
Karten,
2013).
Similarly,
long
descending
telencephalic
efferents
in

chickens
contribute
to
the
recurrent
axonal
connections
within
the
column,
akin
to
layers

5b
and
6
of
the
mammalian
neocortex.
The
avian
visual
wulst
also
has
circuitry
and

physiological
properties
that
are
similar
to
the
mammalian
visual
cortex
(Karten,
1969;

Karten,
2013).
For
example,
like
the
cat
and
monkey
visual
cortex,
the
visual
wulst
includes

precise
retinotopic
organization,
selectivity
for
orientation,
and
selectivity
for
direction
of

movement
(Pettigrew
&
Konishi,
1976).
Together,
these
studies
indicate
that
birds
and

mammals
use
homologous
neural
circuits
to
process
visual
information.
Thus,
controlled-‐
rearing
experiments
with
chicks
can
be
used
to
inform
the
development
of
vision
in

humans.

11

Finally,
while
chickens
have
less
advanced
visual
systems
than
humans,
this
should

not
be
seen
as
a
problem.
When
attempting
to
understand
a
particular
phenomenon,
it
is

often
valuable
to
use
the
simplest
system
that
demonstrates
the
properties
of
interest.

Pioneering
research
in
neuroscience
and
genetics
has
relied
heavily
on
this
strategy—for

example,
researchers
have
used
Aplysia
to
study
the
physiological
basis
of
memory
storage

in
neurons
(e.g.,
Kandel,
2007),
C.
Elegans
to
study
the
mechanisms
of
molecular
and

developmental
biology
(e.g.,
Brenner,
1974),
and
Drosophila
to
study
the
mechanisms
of

genetics
(e.g.,
Bellen,
Tong,
&
Tsuda,
2010).
In
a
similar
vein,
the
study
of
newborn
chicks

can
offer
an
important
window
onto
the
emergence
of
high-‐level
visual
abilities
like

invariant
object
recognition.

Using
automated
controlled
rearing
to
explore
the
origins
of
object
perception

Historically,
newborn
subjects’
behavior
has
been
quantified
through
direct

observation
by
trained
researchers.
While
direct
observation
has
revealed
many
important

insights
about
human
development,
this
approach
has
limitations:
researchers
can
only

observe
a
small
number
of
subjects
simultaneously
and
there
are
constraints
on
the
extent

and
resolution
of
these
observations.

Recent
technological
advances
in
automated
image-‐based
tracking
provide
a

solution
to
these
limitations
by
allowing
researchers
to
collect
large
amounts
of
precise
and

accurate
behavioral
data
(Dell
et
al.,
2014).
Further,
image-‐based
tracking
uses
a
digital

recording
of
the
animal’s
behavior,
which
maintains
an
objective
view
of
events.
This

increases
the
repeatability
of
analyses,
while
allowing
subjects
to
be
tracked
with
high

spatiotemporal
resolution.
Finally,
and
perhaps
most
importantly,
automated
approaches

12

eliminate
the
possibility
of
experimenter
bias
(e.g.,
bias
that
may
occur
when
coding
the

subject’s
behavior,
presenting
stimuli
to
the
subject,
or
deciding
whether
to
include
the

subject
in
the
final
analysis).

The
automated
controlled-‐rearing
method
allows
researchers
to
raise
newborn

chicks
for
several
weeks
within
controlled-‐rearing
chambers
(for
details
see
Wood,
2013).

The
chambers
track
and
record
all
of
the
chicks’
behavior
(9
samples/second,
24

hours/day,
7
days/week),
providing
a
complete
digital
record
of
each
subject’s
behavior

across
their
lifespan.
This
technique
produces
hundreds
of
hours
of
data
for
each
subject,

allowing
researchers
to
measure
chicks’
emerging
visual-‐cognitive
abilities
with
high

precision.

For
example,
the
studies
in
this
dissertation
contain
a
combined
25,200
hours
of

test
data.
In
contrast,
previous
manual
testing
approaches
have
typically
collected
only
5-‐
10
minutes
of
test
data
from
each
newborn
subject
(e.g.,
Martinho
&
Kacelnik,
2016;

Mascalzoni,
Regolin,
&
Vallortigara,
2010;
Regolin,
Rugani,
Stancher,
&
Vallortigara,
2011;

Regolin
&
Vallortigara,
1995;
Rosa-‐Salva,
Grassi,
Lorenzi,
Regolin,
&
Vallortigara,
2016;

Rosa-‐Salva,
Regolin,
&
Vallortigara,
2010;
Vallortigara,
Regolin,
&
Marconato,
2005).
The

smaller
amount
of
data
collection
in
manual
approaches
does
not
permit
analyses
of

individual
subjects’
performance
levels.
It
also
generates
greater
variability
(i.e.,
standard

deviation
sizes)
in
the
collected
data,
which
can
produce
false
positives
and
obscure
true

effects.

Importantly,
the
controlled-‐rearing
chambers
also
make
it
possible
to
control
all
of

the
chicks’
visual
object
experiences.
The
chambers
contain
no
real-‐world
(solid,
bounded)

objects,
and
object
stimuli
are
presented
to
the
chick
by
projecting
virtual
objects
onto
two

13

display
walls
situated
on
opposite
sides
of
the
chamber.
Thus,
the
chicks’
visual
object

experiences
are
limited
to
the
virtual
objects
presented
on
the
display
walls.

Research
using
this
automated
controlled-‐rearing
method
with
newborn
chicks
has

revealed
impressive
visuo-‐cognitive
abilities
at
the
onset
of
vision.
Newborn
chicks
that
are

raised
with
a
single
object
seen
from
a
limited
range
of
viewpoints
are
able
to
perform

viewpoint
invariant
recognition,
identifying
the
object
across
large
viewpoint
changes

(Wood,
2013,
2015).
Once
a
chick
has
formed
a
representation
of
an
object,
that
object
can

be
recognized
rapidly
(Wood
&
Wood,
2016b).
Moreover,
newborn
chicks
are
able
to
solve

the
visual
binding
problem,
binding
together
color
and
shape
features
into
integrated

representations
(Wood,
2014).

Are
specific
visual
experiences
necessary
for
the
development
of
these
robust

representations?
Further
research
using
this
automated
controlled-‐rearing
method
has

revealed
a
number
of
constraints
on
newborns’
abilities
to
build
object
representations.

For
example,
newborn
chicks
require
visual
experience
with
slowly
changing
objects
in

order
to
build
invariant
object
representations
(Wood
&
Wood,
2016a).
In
Wood
&
Wood

(2016)
we
raised
chicks
with
a
single
object
that
rotated
at
either
a
fast,
medium,
or
slow

speed.
Then
we
tested
(1)
whether
chicks
could
recognize
the
object
when
shown
from
a

novel
viewpoint
and
(2)
whether
chicks
preferred
the
original
object
shown
from
the

original
viewpoint
range
or
the
original
object
shown
from
a
novel
viewpoint
range.
We

found
that
when
newborn
chicks
are
raised
with
an
object
that
moves
quickly,
they
build

view-‐specific
representations
that
fail
to
generalize
to
novel
viewpoints
and
rotation

speeds.
Conversely,
when
newborn
chicks
are
raised
with
a
slowly
moving
object,
they

14

build
representations
that
are
specific
to
the
identity
of
the
object
and
tolerant
to
changes

in
viewpoint
and
rotation
speed.

Additionally,
newborn
chicks
require
visual
experience
with
smoothly
changing

objects
to
build
object
representations
(Wood,
2016;
Wood,
Prasad,
Goldman,
&
Wood,

2016;
Wood
&
Wood,
under
review).
In
Wood
&
Wood
(under
review),
we
raised
chicks

with
either
a
smoothly
rotating
object
or
a
non-‐smoothly
rotating
object
(i.e.,
the
same

animation
frames
of
the
object,
but
with
the
order
of
the
frames
scrambled).
As
in
Wood
&

Wood
(2016),
we
tested
(1)
whether
the
chicks
recognized
the
object
when
presented
from

novel
viewpoints
and
(2)
whether
the
chicks
preferred
the
original
object
shown
from
the

original
viewpoint
range
over
the
original
object
shown
from
a
novel
viewpoint
range.

When
chicks
were
raised
with
an
object
that
moved
non-‐smoothly,
they
built

representations
that
were
less
view-‐invariant
(i.e.,
less
selective
for
the
identity
of
the

object
and
less
tolerant
to
changes
in
viewpoint).
Similarly,
in
Wood
et.
al
(2016),
we
raised

chicks
with
either
natural
(smooth)
or
unnatural
(non-‐smooth)
sequences
of
images.
Both

sequences
showed
the
same
images,
but
the
natural
sequences
showed
different

viewpoints
of
the
same
object
and
the
unnatural
sequences
showed
different
images
of

different
objects.
While
the
chicks
that
were
raised
with
natural
sequences
were
able
to

recognize
the
images
from
the
familiar
sequence,
the
chicks
that
were
raised
with
the

unnatural
sequences
could
not
recognize
the
familiar
images.
Thus,
the
newborn
visual

system
is
calibrated
to
operate
over
natural
(smooth)
visual
input.
Overall,
these
prior

studies
demonstrate
that
newborns
require
specific
types
of
visual
inputs
to
learn
invariant

representations
of
objects.

15

The
goal
of
the
studies
presented
in
this
dissertation
is
to
explore
(1)
what
object

recognition
abilities
are
present
at
the
onset
of
vision
and
(2)
what
visual
experiences
are

necessary
to
develop
these
abilities.
The
studies
provide
evidence
that
newborn
animals

can
build
abstract
object
representations
that
generalize
across
novel
viewing
situations.

Thus,
newborn
visual
systems
are
capable
of
impressive
visual
recognition
abilities
at
the

onset
of
vision.
However
these
abilities
do
not
develop
automatically.
Instead,
specific

types
of
visual
experiences
are
necessary
for
newborn
chicks
to
learn
how
to
recognize

objects.
Therefore,
I
also
present
evidence
for
developmental
constraints
on
object
and
face

recognition.

Summary
of
the
Current
Studies

Chapter
2:
Newborn
chicks
segment
objects
from
backgrounds
at
the
onset
of
vision

The
ability
to
segment
objects
from
backgrounds
is
critical
for
real-‐world
object

recognition.
Although
previous
research
has
examined
object
segmentation
in
human

infants
and
patients
recovering
from
blindness,
we
still
do
not
know
whether
the
newborn

brain
is
able
to
segment
swaths
of
retinal
input
into
objects
with
meaningful
boundaries.

One
possibility
is
that
the
development
of
object
segmentation
is
protracted,
requiring

extensive
visual
experiences
with
objects
and
scenes.
Another
possibility
is
that
the

newborn
visual
system
can
begin
segmenting
objects
from
backgrounds
at
the
onset
of

visual
experience.
In
Chapter
2,
we
used
an
automated
controlled-‐rearing
method
to

investigate
whether
newborn
chicks
can
segment
the
first
object
they
see
in
their
life.
We

presented
chicks
with
a
single
object
rotating
on
a
single
background,
and
tested
whether

the
chicks
could
recognize
that
object
when
it
was
presented
on
novel
backgrounds
from

16

novel
viewpoints.
Our
results
indicate
that
the
ability
to
segment
objects
from
the

background
is
present
at
the
onset
of
vision.

Chapter
3:
The
development
of
background-‐invariant
object
recognition
in
visually
naïve

animals

Chapter
2
demonstrates
that
the
newborn
visual
system
is
able
to
segment
objects

without
extensive
visual
experience.
What
experiential
factors
drive
this
early
emerging

ability?
Prior
research
on
human
infants
and
patients
recovering
from
blindness
suggests

that
object
motion
is
necessary
and
sufficient
for
learning
how
to
segment
objects.
In

Chapter
3,
we
tested
this
hypothesis
directly
by
examining
whether
motion
is
sufficient
for

building
background-‐invariant
object
representations
at
the
onset
of
vision.
Specifically,
we

tested
whether
chicks
require
visual
experience
with
objects
moving
on
backgrounds
in

order
to
develop
background-‐invariant
object
recognition.
Surprisingly,
we
found
that

newborn
chicks
that
were
raised
with
an
object
moving
across
no
background
scene
(a

homogenous
white
background)
were
impaired
at
recognizing
the
object
on
background

scenes
(relative
to
chicks
raised
with
a
single
object
moving
on
background
scenes).
Thus,

visual
experience
with
objects
moving
on
backgrounds
is
required
to
learn
background-‐
invariant
recognition.
This
result
provides
evidence
for
a
novel
constraint
on
the

development
of
background-‐invariant
recognition.

17

Chapter
4:
Newborn
chicks
generate
view-‐invariant
object
representations
from
sparse
visual

input

Previous
research
has
demonstrated
that
chicks
can
build
a
view-‐invariant

representation
of
the
first
object
seen
in
life.
However,
these
studies
do
not
reveal
the

amount
of
visual
input
that
is
needed
to
build
these
abstract
representations.
Chapter
2

demonstrated
that
visual
input
of
a
single
object
moving
along
a
single
background
is

sufficient
to
build
a
view-‐invariant
representation.
In
Chapter
4,
we
investigated
whether

extremely
sparse
visual
input—only
3
frames
of
object
motion—is
sufficient
for
the

newborn
brain
to
build
a
view-‐invariant
representation
of
the
first
object
seen
in
life.

The

results
demonstrate
that
chicks
raised
with
only
3
frames
of
object
motion
can
build
view-‐
invariant
object
representations.
Notably,
however,
performance
was
not
fully
invariant

across
all
test
viewpoints.
Chicks’
performance
varied
as
a
function
of
the
amount
of
self-‐
occlusion
in
a
given
test
viewpoint
of
the
object,
but
not
by
lower-‐level
similarities
between

the
test
images
and
imprinted
images
(i.e.,
pixel-‐level
and
V1-‐level
neuronal

representations).
Taken
together,
Chapters
2
and
4
indicate
that
newborn
visual
systems

are
highly
generative,
building
representations
that
allow
the
outputs
of
object
recognition

to
generalize
beyond
visual
input
acquired
by
the
subject.

Chapter
5:
Face
recognition
in
newborn
chicks
at
the
onset
of
vision

Thus
far,
research
on
visual
recognition
in
newborn
chicks
has
focused
on
basic-‐
level
object
recognition.
Chapters
2-‐4
demonstrate
that
newborn
chicks
can
build
invariant

object
representations
that
differ
at
a
basic-‐level,
i.e.,
objects
that
have
different
features

and
overall
configurations.
However,
the
origins
of
subordinate-‐level
recognition
remain

18

largely
unknown.
Objects
that
vary
at
the
subordinate
level
share
a
general
configuration

and
set
of
features.
Therefore,
subordinate-‐level
recognition
requires
more
fine-‐grained

discrimination
than
basic-‐level
object
recognition.
In
Chapter
5,
we
use
faces
as
a

prototypical
example
of
subordinate-‐level
objects.
Given
that
this
is
the
first
study
to
test

face
recognition
in
newborn
chicks,
Chapter
5
was
primarily
exploratory.
We
tested
chicks’

recognition
abilities
across
a
wide
range
of
face
differences
and
found
that
newborn
chicks

are
sensitive
to
changes
in
a
face’s
age,
gender,
and
orientation
(upright
versus
inverted).

Chapter
6:
A
slowness
constraint
on
the
development
of
view-‐invariant
face
recognition

The
ability
to
recognize
faces
in
novel
viewing
situations
is
critical
for
everyday

social
interaction.
Invariant
face
recognition
is
computationally
complex
because
two
faces

can
look
extremely
similar
from
matching
viewpoints
(since
all
faces
share
the
same

general
configuration
of
features),
but
the
same
face
can
provide
vastly
different
retinal

input
when
viewed
from
different
viewpoints.
The
results
of
Chapter
5
indicate
that

newborn
chicks
are
able
to
recognize
faces
at
the
onset
of
vision.
In
Chapter
6,
we
build

upon
this
finding
by
testing
newborns’
ability
to
build
view-‐invariant
representations
of

faces.

Our
results
demonstrate
the
newborn
chicks
are
able
to
build
a
view-‐invariant

representation
of
the
first
face
they
see
in
their
life.
After
establishing
that
newborn
chicks

are
capable
of
view-‐invariant
face
recognition,
we
then
tested
whether
the
development
of

this
ability
is
subject
to
the
same
constraints
as
object
recognition.
In
previous
work,
we

showed
that
the
development
of
object
recognition
requires
visual
experience
with
slowly

moving
objects.
To
examine
whether
this
“slowness
constraint”
also
applies
to
the

19

development
of
face
recognition,
we
systematically
manipulated
the
speed
of
the
virtual

face
presented
to
the
chicks.
As
with
basic-‐level
objects,
we
found
that
the
speed
of
face

motion
during
encoding
directly
affected
the
amount
of
identity
information
and
viewpoint

information
contained
in
each
newborn
face
representation.
Thus,
newborns’
face
and

object
representations
are
subject
to
the
same
“slowness
constraint.”
These
results
suggest

that
newborn
chicks’
representations
of
objects
and
faces
may
rely
on
some
shared
sets
of

computations.

20

Chapter
2:
Newborn
chicks
segment
objects
from
backgrounds
at
the
onset
of
vision

Abstract

To
perceive
the
world
successfully,
the
visual
system
must
learn
to
recognize
objects
across

novel
backgrounds.
To
date,
however,
the
development
of
this
ability
is
poorly
understood.

While
previous
studies
have
shown
that
newborn
animals
can
recognize
objects
presented

on
homogenous
backgrounds,
it
is
unknown
whether
newborns
can
segment
objects
from

complex
backgrounds
and
recognize
those
objects
across
novel
viewing
situations.
To

address
this
issue,
we
raised
newborn
chicks
in
strictly
controlled
environments
that

contained
a
single
virtual
object
moving
on
a
single
background.
We
then
used
an

automated
testing
procedure
to
examine
whether
the
chicks
could
recognize
that
object

across
novel
backgrounds
and
novel
viewpoints.
Despite
receiving
experience
with
just
a

single
object
moving
on
a
single
background,
the
majority
of
chicks
developed
robust
view-‐
invariant
and
background-‐invariant
object
recognition
abilities.
These
results
show
that

advanced
object
recognition
abilities
can
develop
rapidly
in
newborn
brains
from
sparse

visual
input
about
objects.

21

Introduction

To
recognize
objects
successfully,
individuals
must
perform
a
difficult
task:
they

must
segment
objects
from
complex
background
scenes
and
build
abstract
representations

that
generalize
to
novel
viewing
situations.
While
this
task
feels
effortless
to
human
adults,

it
poses
a
major
computational
challenge
(Pinto,
Cox,
&
DiCarlo,
2008).
The
retinal
image
of

an
object
can
change
radically
when
the
object
is
presented
on
different
backgrounds.

Thus,
the
visual
system
must
build
“background-‐invariant”
object
representations
that
are

selective
for
object
identity
and
tolerant
to
background
changes
and
other
identity-‐
preserving
image
transformations
(e.g.,
changes
in
viewpoint).
These
background-‐invariant

representations
are
abstract
insofar
as
they
imply
knowledge
beyond
physical
similarities

between
images
of
an
object.
What
are
the
origins
of
this
foundational
visual
ability?
Can

newborn
brains
begin
building
background-‐invariant
object
representations
at
the
onset
of

vision,
or
does
this
ability
emerge
gradually
over
development?

Studies
of
blind
individuals
who
recover
sight
following
surgery
have
provided

important
insights
into
how
the
visual
system
learns
to
recognize
objects
(McKyton,
Ben-‐
Zion,
Doron,
&
Zohary,
2015;
Ostrovsky
et
al.,
2009).
For
example,
newly-‐sighted

individuals
use
luminance,
hue,
and
motion
cues
to
segment
and
recognize
objects.
Studies

of
infants
also
suggest
that
motion
cues
play
a
key
role
in
the
development
of
object

perception
(Spelke,
1990;
Xu,
2007).
However,
neither
studies
of
infants
nor
studies
of

blind
individuals
who
recover
sight
can
fully
reveal
the
experiential
factors
that
enable

object
recognition
to
emerge
in
the
brain.
Human
infants
cannot
ethically
be
raised
in

controlled
environments
from
birth,
so
it
is
not
possible
to
systematically
manipulate
their

visual
experiences.
Even
infants
who
are
just
a
few
months
old
have
already
acquired

22

hundreds
of
hours
of
experience
with
a
natural
visual
world.
Thus,
with
human
infants,
it
is

not
possible
to
distinguish
whether
early
emerging
abilities
are
innate
or
learned
from

postnatal
visual
experience.
Similarly,
when
a
blind
patient
recovers
sight
following

surgery,
the
patient
is
immediately
confronted
with
a
rich
visual
world.
As
a
result,
it
is
not

possible
to
isolate
the
specific
visual
experiences
that
are
necessary
to
develop
object

recognition.
Additionally,
in
studies
of
sight-‐restored
patients,
visual
deprivation
at
birth

leads
to
cross-‐modal
reorganization
that
makes
the
visual
cortex
of
newly-‐sighted
patients

significantly
different
from
a
newborn
visual
cortex
(Collignon
et
al.,
2015;
Maidenbaum
et

al.,
2014).

Here,
we
offer
a
complementary
approach
for
studying
the
origins
of
object

recognition:
controlled-‐rearing
experiments
with
newborn
animals.
In
controlled-‐rearing

experiments,
it
is
possible
to
systematically
manipulate
the
visual
experiences
provided
to

newborn
subjects
and
measure
the
effects
of
those
manipulations
on
the
development
of

object
recognition
(Wood
&
Wood,
2016a).
Controlled-‐rearing
experiments
therefore

provide
an
experimental
avenue
for
probing
how
object
recognition
emerges
in
newborn

brains
as
a
function
of
specific
visual
experiences.
In
the
present
study,
we
used
an

automated
controlled-‐rearing
method
to
examine
whether
newborn
animals
(domestic

chicks)
are
capable
of
background-‐invariant
object
recognition
at
the
onset
of
vision.

Controlled-‐rearing
experiments
of
object
recognition
in
newborn
chicks

We
used
newborn
chicks
(Gallus
gallus)
as
an
animal
model
because
they
are
an

ideal
model
system
for
studying
the
origins
of
object
recognition.
First,
newborn
chicks
can

perform
advanced
object
recognition
tasks.
For
instance,
newborn
chicks
are
capable
of

23

view-‐invariant
object
recognition
at
the
onset
of
vision
(Wood,
2013;
Wood
&
Wood,

2015a)
and
can
build
integrated
object
representations
with
bound
color-‐shape
features

(Wood,
2014).
The
present
study
extends
this
work
by
examining
whether
newborn
chicks

are
capable
of
background-‐invariant
object
recognition.
This
is
an
important
extension
to

the
literature
because
background-‐invariant
object
recognition
is
necessary
for
recognizing

objects
in
natural
visual
environments.
Second,
newborn
chicks
are
highly
precocial
and

can
be
raised
in
strictly
controlled
environments
immediately
after
hatching
(Horn,
2004;

Vallortigara,
2012;
Wood,
2013).
As
a
result,
it
is
possible
to
examine
the
specific
visual

inputs
that
cause
object
recognition
to
emerge
in
newborn
brains
(Wood,
2016;
Wood
et

al.,
2016;
Wood
&
Wood,
2016a).
Third,
chicks
imprint
to
objects
seen
soon
after
hatching,

which
provides
a
natural
behavioral
response
that
can
be
used
to
test
their
object

recognition
abilities
without
training
(Horn,
2004).
Finally,
avian
and
mammalian
brains

use
homologous
neural
circuits
to
process
sensory
input
(Karten,
2013).
Since
avian
and

mammalian
brains
contain
common
neural
circuits,
controlled-‐rearing
studies
of
newborn

chicks
can
inform
our
understanding
of
the
development
of
vision
in
humans.

The
present
study

During
the
input
phase
(Days
1-‐5),
we
raised
newborn
chicks
in
strictly
controlled

environments
that
contained
a
single
virtual
object
moving
on
a
single
background
scene.

During
the
test
phase
(Days
6-‐12),
we
tested
whether
the
chicks
could
recognize
that
object

when
it
was
presented
on
the
same
or
novel
background
scenes.
Our
experiments
used

moving
objects
as
stimuli
during
the
input
and
test
phases
because
prior
work
has

demonstrated
that
motion
cues
are
critical
to
object
segmentation
in
infants
(Arterberry
&

24

Yonas,
2000;
Johnson,
2003)
and
newly-‐sighted
patients
(Ostrovsky
et
al.,
2009).
To

preview
the
results,
we
found
that
newborn
chicks
are
capable
of
robust
view-‐invariant

and
background-‐invariant
object
recognition.
In
fact,
recognition
performance
was
equally

good
whether
the
object
was
presented
in
familiar
situations
(familiar
views,
familiar

backgrounds)
and
novel
situations
(novel
views,
novel
backgrounds).
Thus,
invariant

object
recognition
can
develop
rapidly
in
newborn
brains
from
sparse
visual
input
with

objects.

Methods

Subjects

Thirty-‐one
Rhode
Island
Red
chicks
of
unknown
sex
were
tested.
A
minimum

sample
size
was
determined
based
on
prior
studies
(Wood,
2013),
and
the
ultimate
sample

size
tested
in
the
present
study
was
about
three
times
the
minimum
sample
size
to

accommodate
counter-‐balancing
of
the
stimuli.
No
subjects
were
excluded
from
the

analyses.
The
eggs
were
obtained
from
a
local
distributer
and
incubated
in
darkness
in
an

OVA-‐Easy
incubator
(Brinsea
Products
Inc.,
Titusville,
FL).
To
avoid
exposing
the
chicks
to

any
extraneous
visual
input,
we
used
night
vision
goggles
to
move
the
chicks
in
darkness

from
the
incubation
room
to
the
controlled-‐rearing
chambers.
Each
chick
was
raised
within

its
own
chamber.
This
experiment
was
approved
by
The
University
of
Southern
California

Institutional
Animal
Care
and
Use
Committee.

25

Controlled-‐Rearing
Chambers

The
controlled-‐rearing
chambers
(66
cm
length
×
42
cm
width
×
69
cm
height)
were

constructed
from
white,
high-‐density
polyethylene.
The
chambers
were
devoid
of
all
real-‐
world
(solid,
bounded)
objects.
To
present
object
stimuli
to
the
chicks,
virtual
objects
were

projected
on
two
displays
walls
situated
on
opposite
sides
of
the
chamber.
The
display

walls
were
19”
liquid
crystal
display
monitors
(1440
×
900
pixel
resolution).
Food
and

water
were
provided
ad
libidum
in
transparent
troughs
in
the
ground.
We
used
grain
as

food
because
a
heap
of
grain
does
not
behave
like
an
object
(i.e.,
a
heap
of
grain
does
not

maintain
a
rigid,
bounded
shape).
The
floors
were
constructed
from
wire
mesh
supported

by
transparent
beams.
Micro-‐cameras
in
the
ceilings
of
the
chambers
recorded
all
of
the

chicks’
behavior,
and
the
video
feed
was
analyzed
with
automated
image-‐based
tracking

software
(EthoVision
XT,
Noldus
Information
Technology,
Leesburg,
VA).
This
software

calculated
the
amount
of
time
each
chick
spent
within
zones
(22
cm
×
42
cm)
next
to
the

left
and
right
display
walls.
All
of
the
chicks’
behavior
(9
samples/second,
24
hours/day,
7

days/week)
was
tracked
and
recorded
across
the
two-‐week
duration
of
the
experiment.
In

total,
we
collected
8,928
hours
of
video
footage
for
this
experiment
(24
hours
per
day
×
12

days
×
31
subjects).

Procedure

During
the
input
phase
(Days
1-‐5
of
post-‐natal
visual
experience
3
),
the
chicks
were

raised
in
controlled-‐rearing
chambers
that
contained
a
single
virtual
object
rotating
around

3
All chicks were exposed to the imprinting stimuli for 5 days. Due to a power outage, 20 chicks
spent an additional 2 days in the input phase in complete darkness, but performance during the
test phase was not affected by the power outage in the input phase.

26

a
frontoparallel
horizontal
axis
(Figure
1).
The
object
rotated
continuously,
completing
a

full
rotation
every
15
seconds.
The
object
was
presented
on
one
of
three
background

scenes
(Figure
2;
12
chicks
were
imprinted
to
Background
1;
8
chicks
were
imprinted
to

Background
2;
and
11
chicks
were
imprinted
to
Background
3).
The
object
and
background

appeared
on
one
display
wall
at
a
time
and
switched
to
the
opposite
display
wall
every
two

Figure 1. Illustration of a controlled-rearing chamber. The chambers contained no real-
world objects. To present object stimuli to the chick, virtual objects were projected on two
display walls situated on opposite sides of the chamber. (A) During the input phase (Days 1-5
of post-natal visual experience), the chicks were raised with a virtual object rotating on a
single background scene. (B) During the test trials (Days 6-14 of post-natal visual
experience), the chicks were presented with the imprinted object on one display wall and a
novel object on the opposite display wall. Each object rotated in front of either a novel
background scene or the imprinted background scene.

27

hours
(following
a
one-‐minute
period
of
darkness).
The
display
wall
that
was
not
showing

the
imprinted
object
was
white.
Figure
1
illustrates
how
the
imprinting
stimuli
were

presented
on
the
display
walls
during
the
input
phase.

During
the
test
phase
(Days
6-‐12
of
post-‐natal
visual
experience),
we
tested

whether
the
chicks
could
recognize
their
imprinted
object
when
the
object
was
presented

on
familiar
and
novel
backgrounds.
On
each
test
trial,
the
imprinted
object
appeared
on
one

display
wall,
and
a
novel
object
appeared
on
the
opposite
display
wall
(Figure
3).
The
novel

Figure 2. The backgrounds and objects. Each chick was imprinted to one of the two objects
(right) rotating 360° along the elevation axis on one of three backgrounds (left). During the
input phase, the chicks only saw a single object rotating on a single background. The
remaining backgrounds were used as novel backgrounds for the test phase, and the remaining
object was used as the unfamiliar object for the test phase. During the test phase, the
imprinted object was shown on one display wall (on either an imprinted or novel background)
and the unfamiliar object was shown on the opposite display wall (on either an imprinted or
novel background). The objects were presented rotating along the elevation rotation at 0°, 30°,
and 60° azimuth viewpoint changes during the test phase.

28

object
had
a
similar
size,
color,
motion
speed,
and
motion
trajectory
as
the
imprinted

object.
The
two
objects
were
modeled
after
those
used
in
previous
studies
that
tested
for

invariant
object
recognition
in
adult
rats
(Zoccolan
et
al.,
2009)
and
newborn
chicks

(Wood,
2013,
2015).
The
test
objects
were
presented
on
all
possible
combinations
of
the

three
background
scenes
(e.g.,
Background
1
versus
Background
2,
Background
1
versus

Background
3,
Background
2
versus
Background
3,
etc.).
During
test,
the
objects
rotated

60°
around
a
fronto-‐parallel
horizontal
axis
(as
in
the
input
phase).
In
each
test
trial,
the

test
objects
were
shown
from
three
possible
viewing
angles:
0°
change
in
azimuth
rotation,

30°
change
in
azimuth
rotation,
and
60°
change
in
azimuth
rotation.
The
imprinted
object

and
the
unfamiliar
object
were
always
shown
from
the
same
viewing
angle
in
each
test

trial.

We
grouped
the
test
trials
into
four
background
conditions.
In
the
Both
Objects
Old

Background
condition,
both
the
imprinted
object
and
the
unfamiliar
object
were
shown
on

the
familiar
background
from
the
input
phase.
In
the
Imprinted
Object
Old
Background

condition,
the
imprinted
object
was
shown
on
the
familiar
background
from
the
input

phase,
and
the
unfamiliar
object
was
shown
on
one
of
the
two
unfamiliar
backgrounds.
In

the
Unfamiliar
Object
Old
Background
condition,
the
unfamiliar
object
was
shown
on
the

familiar
background
from
the
input
phase,
and
the
imprinted
object
was
shown
on
one
of

the
two
unfamiliar
backgrounds.
In
the
Both
Objects
New
Background
condition,
the

imprinted
object
and
the
unfamiliar
object
were
each
shown
on
an
unfamiliar
background.

This
set
of
conditions
allowed
us
to
distinguish
between
three
hypotheses
regarding

the
origins
of
object
segmentation
and
background-‐invariant
object
recognition:
(H1)

newborn
visual
systems
are
able
to
segment
objects
from
backgrounds
and
recognize

29

objects
across
novel
backgrounds;
(H2)
newborn
visual
systems
are
able
to
segment

objects
from
backgrounds,
but
there
is
a
cost
for
recognizing
objects
across
novel

backgrounds
compared
to
familiar
backgrounds;
(H3)
newborn
visual
systems
are

incapable
of
segmenting
objects
from
backgrounds.

The
chicks
received
24
test
trials
per
day
at
the
rate
of
one
trial
per
hour.
Each
test

trial
lasted
40
minutes,
and
was
followed
by
a
20-‐minute
rest
period.
During
the
rest

Figure 3. The experimental design. This schematic shows how the virtual objects were
presented on the two display walls, during sample four-hour periods. During the input phase,
the chicks were exposed to a single virtual object rotating on a single background image.
During the test phase, we measured the chicks’ object recognition performance when the
imprinted object was presented on the familiar background (“Both Objects Old Background”
and “Imprinted Object Old Background” conditions) and novel backgrounds (“Both Objects
New Background” and “Unfamiliar Object Old Background” conditions). The object was also
presented from familiar viewpoints (0° azimuth rotation) and novel viewpoints (30° & 60°
azimuth rotations) on different test trials.

30

periods,
the
animation
from
the
input
phase
appeared
on
one
display
wall,
and
a
white

screen
appeared
on
the
other
display
wall.
14
of
the
chicks
were
imprinted
to
Object
1,

with
Object
2
serving
as
the
novel
object,
and
17
of
the
chicks
were
imprinted
to
Object
2,

with
Object
1
serving
as
the
novel
object.

Results

Overall
Recognition
Performance

The
results
are
shown
in
Figure
4.
Performance
was
well
above
chance
level
(50%)

in
each
background
condition
(one
sample
t-‐tests,
all
Ps
<
10
-‐7
).
Similarly,
performance
was

also
well
above
chance
in
each
viewpoint
condition
(one
sample
t-‐tests,
all
Ps
<
10
-‐10
).

Thus,
the
chicks
were
able
to
recognize
their
imprinted
object
whether
it
was
presented
on

familiar
or
novel
backgrounds
and
whether
the
object
was
presented
from
familiar
or
novel

viewpoints.

To
examine
whether
performance
differed
across
background
and
viewpoint

conditions,
we
performed
a
repeated-‐measures
ANOVA
with
within-‐subjects
factors
of

Background
Condition
and
Viewpoint
Angle.
The
ANOVA
revealed
a
significant
main
effect

of
Background
Condition
(Greenhouse-‐Geisser
adjusted,
F(1.778,
53.354)
=
9.455,
p
=

.0005)
and
a
significant
interaction
between
Background
Condition
and
Viewpoint
Angle

(F(6,
180)
=
2.509,
p
=
.023).
The
main
effect
of
Viewpoint
Angle
was
not
significant
(F(2,

60)
=
.032,
p
=
.969).
The
significant
interaction
between
Background
Condition
and

Viewpoint
Angle
appeared
to
be
driven
by
low
performance
when
the
viewpoint
angle
was

0°
and
the
imprinted
object
was
presented
on
the
familiar
background
(with
the
unfamiliar

object
presented
on
the
novel
background).
Although
the
test
animation
of
the
imprinted

object
was
identical
to
the
imprinting
animation
shown
during
the
input
phase,

31

performance
in
this
viewpoint-‐background
combination
was
significantly
lower
than

nearly
all
of
the
other
viewpoint-‐background
combinations
(post
hoc
paired
t-‐tests,
10
out

of
11
Ps
<
.01,
and
all
10
significant
tests
survive
Holm-‐Bonferroni
correction
for
multiple

comparisons).
Correspondingly,
performance
was
significantly
lower
in
the
Imprinted

Object
Old
Background
condition
overall
than
in
all
other
background
conditions
(paired
t-‐
tests,
all
Ps
<
.01,
and
all
three
tests
survive
Holm-‐Bonferroni
correction
for
multiple

Figure 4. Results. The top graph shows the percent of time the chicks spent with the
imprinted object versus novel object for each background condition and viewpoint change.
The dashed line indicates chance performance. Error bars show ±1 standard error (SEM). The
chicks successfully recognized their imprinted object across all background changes and
viewpoint changes. The bottom graphs show the percent of time the chicks spent with the
imprinted object versus the novel object for each (left) background condition and (right)
viewpoint change.

32

comparisons).
There
were
no
significant
differences
between
any
of
the
other
background

conditions
(paired
t-‐tests,
all
Ps
>
.10).

Analysis
of
Change
in
Performance
Over
Time

Overall
performance
by
test
day
is
shown
in
Figure
5.
To
determine
whether

performance
changed
significantly
over
the
course
of
the
test
phase,
we
performed
a

repeated-‐measures
ANOVA
with
the
within-‐subjects
effect
of
Test
Day.
The
ANOVA

revealed
a
significant
main
effect
of
Test
Day
(F(6,
180)
=
8.585,
p
<
10
-‐7
).
A
post-‐hoc

correlation
between
the
Test
Day
and
the
average
performance
for
that
day
revealed
a

significant
positive
relationship
between
Test
Day
and
performance
(r
=
.892,
p
=
.007).

Critically,
however,
performance
was
significantly
above
chance
levels
on
Day
1
(one-‐
sample
t-‐test,
t(30)
=
6.987,
p
<
10
-‐7
).
In
fact,
when
the
analysis
only
included
the
test
trials

Figure 5. Change Over Time. The graph shows the percent of time the chicks spent with the
imprinted object versus novel object for each day of testing. The dashed line indicates chance
performance. The blue shaded ribbon shows ±1 standard error (SEM). While performance
improved throughout the test phase, chicks were able to successfully recognize the imprinted
object on all test days.

33

in
which
the
imprinted
object
was
shown
on
novel
backgrounds,
performance
was
still

significantly
above
chance
levels
on
Day
1
(one-‐sample
t-‐test,
t(30)
=
6.736,
p
<
10
-‐6
).

Analysis
of
Individual
Subject
Performance

Since
we
collected
a
large
number
of
test
trials
from
each
subject,
we
were
able
to

analyze
each
newborn
chick’s
object
recognition
abilities
with
high
precision.
We

computed
each
chick’s
performance
on
each
test
trial
(Figure
6).
Collapsing
across
all
test

trials,
all
of
the
chicks
spent
more
time
with
the
imprinted
object
than
with
the
unfamiliar

object
(one-‐sample
t-‐tests,
all
Ps
<
.03;
all
Ps
survive
Holm-‐Bonferroni
correction
for

Figure 6. Individual Subject Performance. The graphs show the percent of time each chick
spent with the imprinted object versus the novel object. Each chick is represented by a single
marker, with error bars around each subject showing ±1 standard error (SEM). The blue boxes
indicate the 1
st
to 2
nd
quartile and the 2
nd
to 3
rd
quartile of performance, respectively. The
dashed line indicates chance performance. The left graph shows performance across all of the
test conditions, while the right graph shows performance only for the test conditions in which
the imprinted object was shown on a novel background.

34

multiple
comparisons).
After
limiting
the
analysis
to
the
test
trials
in
which
the
imprinted

object
was
shown
on
a
novel
background,
28
of
the
31
chicks
spent
more
time
with
the

imprinted
object
than
with
the
unfamiliar
object
(one-‐sample
t-‐tests,
28
Ps
<
.001;
all
28

significant
Ps
survive
Holm-‐Bonferroni
correction
for
multiple
comparisons).

Discussion

We
used
an
automated
controlled-‐rearing
method
to
examine
whether
newborn

animals
can
build
background-‐invariant
and
view-‐invariant
object
representations
at
the

onset
of
vision.
During
the
input
phase,
we
raised
newborn
chicks
in
strictly
controlled

environments
that
contained
a
single
virtual
object
rotating
on
a
single
background
scene.

During
the
test
phase,
we
examined
whether
the
chicks
could
recognize
that
object
across

novel
backgrounds
and
novel
viewpoints.
Our
results
indicate
that
chicks
are
able
to
form
a

background-‐invariant
and
view-‐invariant
representation
of
the
first
object
they
see
in
their

life.
Further,
the
chicks
were
not
impaired
at
recognizing
the
object
on
unfamiliar

backgrounds
or
from
unfamiliar
viewpoints.
Thus,
newborn
chicks
rapidly
develop
robust

abilities
for
object
segmentation
and
invariant
object
recognition.

It
is
important
to
emphasize
that
while
some
researchers
have
argued
that
visual

systems
solve
invariant
object
recognition
tasks
by
building
complex
3D
representations
of

objects
(e.g.,
Biederman,
1987),
newborn
chicks
could
achieve
invariant
object
recognition

by
building
invariant
representations
of
subfeatures
that
are
smaller
or
less
complex
than

the
entire
object.
These
subfeatures
might
correspond
to
only
a
portion
of
the
object,
or
be

sensitive
to
key
2D,
rather
than
3D,
features
(Alemi-‐Neissi
et
al.,
2013).
In
fact,
many

modern
computational
models
of
invariant
object
recognition
in
primates
explicitly
rely
on

such
subfeatures
(Krizhevsky,
Sutskever,
&
Hinton,
2012;
Yamins
et
al.,
2014).
Regardless

35

of
the
specific
nature
of
these
features,
the
present
results
indicate
that
newborn
chicks
can

build
invariant
features
that
are
tolerant
to
changes
in
viewpoint
and
background
features.

These
findings
replicate
previous
work
showing
that
newborn
chicks
are
capable
of

view-‐invariant
object
recognition
(Wood,
2013,
2015;
Wood
&
Wood,
2015a),
and
extend

the
literature
by
showing
that
chicks
are
also
capable
of
background-‐invariant
object

recognition.
Together,
these
studies
suggest
that
newborn
visual
systems
can
be

surprisingly
powerful:
newborn
chicks
can
build
invariant
object
representations
from

limited
visual
experience
with
objects.
Controlled-‐rearing
studies
of
newborn
chicks

therefore
offer
a
promising
experimental
avenue
for
probing
how
high-‐level
vision

develops
in
the
newborn
brain.

Furthermore,
controlled-‐rearing
studies
offer
an
important
complementary

approach
to
studying
infants
and
newly-‐sighted
patients.
The
present
results
are
consistent

with
findings
that
infants
and
blind
individuals
who
recover
sight
following
medical

intervention
are
able
to
determine
object
boundaries
by
using
motion
cues
(Kellman
et
al.,

1986;
Ostrovsky
et
al.,
2009).
However,
neither
of
these
populations
can
be
used
to

determine
whether
the
visual
system
is
capable
of
object
segmentation
and
recognition
in

the
absence
of
experience
with
a
natural
visual
world.
Our
results
provide
evidence
that

newborn
visual
systems
are
capable
of
segmenting
objects
from
complex
backgrounds
and

building
abstract
representations
of
those
objects.

In
conclusion,
this
study
demonstrates
that
newborn
chicks
have
advanced
visual

processing
machinery.
Newborn
chicks
can
build
background-‐invariant
and
view-‐invariant

object
representations
after
acquiring
visual
experience
with
just
a
single
object
moving
on

a
single
background.
From
a
computer
vision
perspective,
this
is
an
impressive

36

computational
feat
(Pinto
et
al.,
2008).
Modern
machine
learning
techniques
typically

require
thousands
of
training
images
in
order
to
perform
background-‐invariant
and
view-‐
invariant
recognition.
Conversely,
our
findings
demonstrate
that
newborn
visual
systems

can
build
background-‐invariant
and
view-‐invariant
representations
from
sparse
visual

input.
Thus,
controlled-‐rearing
studies
of
newborns
can
provide
important
benchmarks
for

building
computational
models
that
emulate
the
biological
development
of
vision.

37

Chapter
3:
The
development
of
background-‐invariant
object
recognition
in
visually
naïve

animals

Abstract

To
perceive
objects
successfully,
the
visual
system
segments
swaths
of
pigmentation
into

discrete
entities.
The
visual
system
must
segment
objects
from
backgrounds
and
build

abstract
representations
that
generalize
across
novel
viewing
situations.
How
does
this

ability—known
as
“background-‐invariant
object
recognition”—develop
in
newborn
brains?

While
prior
studies
have
demonstrated
that
object
motion
is
necessary
for
learning
how
to

segment
objects
from
backgrounds,
it
is
unknown
whether
object
motion
is
sufficient
for

the
development
of
this
ability.
To
address
this
issue,
we
raised
newborn
chicks
in
strictly

controlled
environments
that
contained
a
single
virtual
object
moving
either
on
no

background
or
on
natural
backgrounds.
We
then
used
an
automated
testing
procedure
to

examine
whether
the
chicks
could
recognize
that
object
across
novel
backgrounds
and

novel
viewpoints.
We
found
that
chicks
raised
without
experience
of
objects
moving
on

backgrounds
showed
impaired
background-‐invariant
object
recognition.
Moreover,
the

chicks’
recognition
performance
improved
throughout
the
test
phase
as
they
acquired

more
experience
with
objects
moving
on
backgrounds.
These
results
suggest
that
object

motion
is
insufficient
for
the
development
of
background-‐invariant
recognition.
The

development
of
this
ability
requires
visual
experience
with
objects
moving
on
backgrounds.

38

Introduction

How
does
object
perception
emerge
in
newborn
brains?
Despite
significant
interest

in
the
origins
of
this
ability,
methodological
barriers
have
largely
prevented
precise

empirical
studies
of
object
perception
in
newborn
humans.
As
a
result,
little
is
known
about

the
role
of
visual
experience
in
the
development
of
this
ability.
In
contrast
to
studies
of

human
infants,
controlled-‐rearing
studies
of
newborn
chicks
are
uniquely
situated
for

examining
the
origins
of
object
perception.
Unlike
human
infants,
chicks
can
be
raised
from

birth
in
strictly
controlled
visual
worlds.
Thus,
researchers
can
systematically
manipulate

the
visual
experiences
provided
to
newborn
chicks
and
examine
which
abilities
emerge

from
those
experiences.
This
approach
makes
it
possible
to
reveal
the
role
of
experience
in

the
development
of
perception
and
cognition.
Recent
developments
in
automated

controlled
rearing
also
provide
an
opportunity
to
probe
the
initial
state
of
object

perception
with
an
unprecedented
degree
of
precision
(Wood,
2013;
Wood
&
Wood,

2015a).

Human
adults
can
parse
natural
visual
scenes
with
relative
ease;
yet,
the
ability
to

translate
patches
of
different
hues
and
luminance
into
unified,
meaningful
object

representations
is
incredibly
challenging
from
a
computational
perspective.
For
example,

to
recognize
an
object
in
a
natural
setting,
the
visual
system
must
solve
at
least
two

abstract
problems.
First,
the
visual
system
must
group
regions
of
different
color
and

luminance
into
individual
objects.
Second,
the
visual
system
must
then
be
able
to
recognize

those
objects
across
large
changes
in
visual
appearance
(for
example,
due
to
changes
in

viewpoint,
lighting,
position,
and
size).
This
ability,
known
as
invariant
object
recognition,

has
been
studied
extensively
in
adult
subjects
(Biederman,
1987;
Logothetis
&
Sheinberg,

39

1996;
Rolls,
2000;
Tanaka,
1996;
Zoccolan,
2015).
However,
little
is
known
about
the

origins
of
visual
parsing
and
invariant
recognition
in
the
newborn
brain.
What
experiences

are
needed
for
the
newborn
visual
system
to
segment
objects
from
backgrounds
and

recognize
those
objects
across
novel
viewing
situations?

Previous
research
examining
the
development
of
object
segmentation
has
generally

focused
on
two
populations:
human
infants
and
adult
patients
recovering
from
blindness.

Results
from
both
of
these
populations
have
converged
on
an
important
finding:
motion

information
is
critical
to
the
early
development
of
object
segmentation
(Arterberry
&

Yonas,
2000;
Johnson,
Schwarzer,
&
Leder,
2003;
Ostrovsky
et
al.,
2009;
Spelke,
1990).

While
these
studies
have
helped
to
elucidate
the
development
of
background-‐invariant

object
recognition,
they
cannot
reveal
how
background-‐invariant
recognition
emerges
in

the
brain.
Although
infants
and
adult
patients
recovering
from
blindness
have
significantly

less
visual
experience
than
sighted
adults
with
normal
visual
systems,
both
populations

have
still
acquired
weeks
to
months
of
natural
visual
experiences
prior
to
testing.

Recently,
automated
controlled-‐rearing
studies
of
newborn
chicks
have
begun
to

tackle
these
questions
by
examining
the
initial
state
of
object
segmentation
and

recognition.
For
instance,
newborn
chicks
are
able
to
build
view-‐invariant
representations

of
the
first
object
seen
in
life
(Wood,
2013,
2015;
Wood
&
Wood,
2015a).
Moreover,
these

studies
provide
converging
evidence
that
motion
information
is
critical
for
the

development
of
object
perception.
In
particular,
newborn
chicks
need
visual
experience

with
objects
that
move
smoothly
and
slowly
over
time
in
order
to
build
view-‐invariant

representations
(Wood,
2016;
Wood
et
al.,
2016;
Wood
&
Wood,
2016a).
Further,
newborn

chicks
can
segment
moving
objects
from
backgrounds
(Chapter
2)
and
fail
to
develop

40

background-‐invariant
recognition
when
objects
are
stationary
(Wood
&
Wood,
in
prep),

consistent
with
prior
work
on
infants
and
patients
recovering
from
blindness.

While
prior
studies
have
demonstrated
that
motion
information
is
necessary
for
the

development
of
object
segmentation
abilities,
it
is
unknown
whether
motion
information
is

also
sufficient.
Can
newborns
build
a
background-‐invariant
object
representation
merely

from
visual
experience
with
a
moving
object
(i.e.,
in
the
absence
of
a
background)?
To

address
this
issue,
we
raised
newborn
chicks
from
birth
with
a
single
moving
object.
In
the

No
Background
condition,
the
chicks
were
raised
with
an
object
moving
on
a
homogenous

white
background.
Thus,
chicks
in
this
condition
had
visual
experience
with
object
motion,

but
did
not
acquire
visual
experience
with
objects
moving
on
patterned
backgrounds
prior

to
testing.
In
the
Background
condition,
chicks
were
raised
with
an
object
moving
on

multiple
backgrounds.

If
motion
information
is
both
necessary
and
sufficient
for
newborns
to
build

background-‐invariant
object
representations,
then
the
chicks
in
the
No
Background

condition
should
successfully
recognize
the
object
across
novel
backgrounds,
despite
never

seeing
the
object
move
across
a
patterned
background.
According
to
this
hypothesis,
object

motion
provides
all
of
the
critical
information
needed
to
segment
objects
and
recognize

objects
across
novel
viewing
conditions.
Conversely,
if
newborns
need
visual
experience

with
objects
moving
over
patterned
backgrounds
to
develop
background-‐invariant

recognition,
then
the
chicks
in
the
No
Background
condition
should
fail
to
recognize
the

object
across
novel
backgrounds.
According
to
this
hypothesis,
while
object
motion
might

be
necessary
for
the
development
of
background-‐invariant
recognition,
it
might
not
be

sufficient.

41

To
preview
the
findings,
our
results
support
the
latter
hypothesis.
Object
motion
is

not
sufficient
for
newborns
to
build
background-‐invariant
representations
of
objects.

Rather,
newborn
chicks
need
visual
experience
with
objects
moving
over
patterned

backgrounds
to
successfully
develop
background-‐invariant
recognition.

Methods

Subjects

Twenty-‐one
Rhode
Island
Red
chicks
of
unknown
sex
were
tested
(11
chicks
in
the

No
Background
condition
and
10
chicks
in
the
Background
condition).
No
subjects
were

excluded
from
the
analyses.
The
sample
size
was
determined
before
the
experiment
was

conducted,
based
on
previous
automated
controlled-‐rearing
experiments
with
newborn

chicks
(Wood,
2013,
2014,
2015).
The
eggs
were
obtained
from
a
local
distributer
and

incubated
in
darkness
in
an
OVA-‐Easy
incubator
(Brinsea
Products
Inc.,
Titusville,
FL).
To

avoid
exposing
the
chicks
to
any
extraneous
visual
input,
we
used
night
vision
goggles
to

move
the
chicks
in
darkness
from
the
incubation
room
to
the
controlled-‐rearing
chambers.

Each
chick
was
raised
within
its
own
chamber.
This
experiment
was
approved
by
The

University
of
Southern
California
Institutional
Animal
Care
and
Use
Committee.

Controlled-‐Rearing
Chambers

Each
subject
was
reared
singly
in
a
controlled-‐rearing
chamber
(66
cm
length
×
42

cm
width
×
69
cm
height)
constructed
from
high-‐density
polyethylene.
The
chambers
were

devoid
of
all
solid,
bounded
objects.
Stimuli
were
presented
on
two
display
walls
situated

on
opposite
sides
of
the
chamber.
The
display
walls
were
19”
liquid
crystal
display

42

monitors
(1440
×
900
pixel
resolution).
Food
and
water
were
provided
ad
libidum
in

transparent
troughs
in
the
ground.
Micro-‐cameras
in
the
ceilings
of
the
chambers
recorded

all
of
the
chicks’
behavior,
and
the
video
feed
was
analyzed
with
automated
image-‐based

tracking
software
(EthoVision
XT,
Noldus
Information
Technology,
Leesburg,
VA).
The

software
calculated
the
amount
of
time
each
chick
spent
within
zones
(22
cm
×
42
cm)
next

to
the
left
and
right
display
walls.
All
of
the
chicks’
behavior
(9
samples/second,
24

hours/day,
7
days/week)
was
tracked
and
recorded
across
the
two-‐week
duration
of
the

experiment.
In
total,
we
collected
7,056
hours
of
video
footage
for
this
experiment
(24

hours
per
day
×
14
days
×
21
subjects).

Procedure

During
the
input
phase
(Days
1-‐5
of
life),
the
chicks
were
raised
in
controlled-‐
rearing
chambers
that
contained
a
single
virtual
object
rotating
around
a
frontoparallel

horizontal
axis
(Figure
7).
The
object
rotated
continuously,
completing
a
full
rotation
every

15
seconds.
In
the
No
Background
condition,
the
object
was
presented
on
a
white,

homogenous
background.
In
the
Background
condition,
the
object
was
presented
on

multiple
backgrounds,
which
changed
every
two
hours.
The
object
appeared
on
one
display

wall
at
a
time
and
switched
to
the
opposite
display
every
two
hours.
In
both
conditions,
the

display
wall
that
was
not
showing
the
imprinted
object
was
white.
Figure
8
illustrates
how

the
stimuli
were
presented
on
the
display
walls
during
the
input
phase.

During
the
test
phase
(Days
6-‐14
of
life),
we
tested
whether
the
chicks
could

recognize
their
imprinted
object
when
the
object
was
presented
on
a
variety
of

backgrounds.
On
each
test
trial,
the
imprinted
object
appeared
on
one
display
wall,
and
a

43

novel
object
appeared
on
the
opposite
display
wall
(Figures
7
&
8).
The
novel
object
had
a

similar
size,
color,
motion
speed,
and
motion
trajectory
as
the
imprinted
object.
The
two

objects
were
modeled
after
those
used
in
previous
studies
that
tested
for
invariant
object

recognition
in
adult
rats
(Zoccolan
et
al.,
2009)
and
newborn
chicks
(Wood,
2013,
2015).

Figure 7. Illustration of a controlled-rearing chamber. The chambers contained no real-
world objects. To present object stimuli to the chick, virtual objects were projected on two
display walls situated on opposite sides of the chamber. (A) During the input phase (Days 1-5
of life), the chicks were raised with a virtual object rotating on no background (No
Background condition) or on multiple backgrounds (Background condition, depicted in A).
(B) During the test trials (Days 6-14 of life), the chicks were presented with the imprinted
object on one display wall and a novel object on the opposite display wall.

44

The
test
objects
were
presented
on
the
same
background
image,
from
the
same
viewpoint

range,
and
rotated
360°
around
a
frontoparallel
horizontal
axis
(as
in
the
input
phase).

In
total,
we
used
24
background
images
during
the
test
phase.
Eight
images
(2

exemplars
×
4
categories)
were
the
images
used
in
the
Background
condition
during
the

input
phase.
Eight
additional
images
(2
exemplars
×
4
categories)
were
from
the
same

categories
as
the
Background
condition
(i.e.,
coasts,
forests,
highways,
and
mountains),
and

Figure 8. The experimental design. This schematic shows how the virtual objects were
presented on the two display walls, during sample four-hour periods. During the input phase,
the chicks were exposed to a single virtual object rotating on a homogenous white background
(A; No Background condition) or eight background images (B; Background condition).
During the test phase (C; both conditions), we measured whether the chicks could recognize
the object across novel backgrounds and novel viewpoints (30° & 60° azimuth rotations) on
different test trials.

45

the
final
eight
images
(2
exemplars
×
4
categories)
were
from
novel
categories
(specifically,

open
countries,
rivers,
deserts,
and
urban
cityscapes).

In
addition
to
changing
the
background
images
across
the
test
trials,
we
also

manipulated
the
viewpoint
of
the
objects
by
rotating
the
test
objects
0°,
30°,
or
60°
along

the
azimuth
axis.
Both
objects
were
shown
from
the
same
viewpoint
range
on
each
test

trial.
We
included
viewpoint
changes
in
the
design
to
examine
whether
newborn
chicks
are

simultaneously
capable
of
both
background-‐invariant
and
view-‐invariant
object

recognition.
Finally,
to
control
for
the
brightness
of
the
objects,
we
equated
the
overall

brightness
of
the
imprinted
object
and
unfamiliar
object
on
the
test
trials
by
decreasing
the

size
of
Object
1
by
11.5%
and
increasing
the
size
of
Object
2
by
11.5%.

The
chicks
received
24
test
trials
per
day
at
the
rate
of
one
trial
per
hour.
Each
test

trial
lasted
40
minutes,
and
was
followed
by
a
20-‐minute
rest
period.
During
the
rest

periods,
the
animation(s)
from
the
input
phase
appeared
on
one
display
wall,
and
a
white

screen
appeared
on
the
other
display
wall.
Ten
of
the
chicks
were
imprinted
to
Object
1,

with
Object
2
serving
as
the
novel
object,
and
11
of
the
chicks
were
imprinted
to
Object
2,

with
Object
1
serving
as
the
novel
object.

Results

Overall
Recognition
Performance

To
compute
object
recognition
performance,
we
measured
the
proportion
of
time

each
chick
spent
with
the
imprinted
object
compared
to
the
novel
object
during
the
test

trials.
Results
are
shown
in
Figure
9.
The
chicks’
overall
recognition
performance
was

significantly
above
chance
levels
in
both
conditions
(one-‐sample
t-‐tests,
No
Background

46

condition:
t(10)
=
4.469,
p
=
.001,
Cohen’s
d
=
1.348;
Background
condition:
t(9)
=
9.292,
p

=
.000007,
Cohen’s
d
=
2.938).
Performance
was
similarly
high
in
the
Background
condition

after
removing
the
test
trials
in
which
the
objects
were
presented
on
familiar
backgrounds

from
the
input
phase
(one-‐sample
t-‐test,
t(9)
=
8.547,
p
=
.00001,
Cohen’s
d
=
2.703).
A

repeated-‐measures
ANOVA
with
the
within-‐subjects
factor
of
Viewpoint
Change
and
the

between-‐subjects
factor
of
Condition
revealed
a
significant
main
effect
of
Condition
(F(1,

19)
=
6.281,
p
=
.021,
η
2

=
.248)
and
a
significant
main
effect
of
Viewpoint
Change

(Greenhouse-‐Geisser
adjusted,
F(1.535,
29.172)
=
6.478,
p
=
.008,
η
2

=
.254).
The

interaction
between
Viewpoint
Change
and
Condition
was
not
significant.

To
further
investigate
the
main
effect
of
Condition,
we
performed
an
independent

samples
t-‐test
comparing
overall
recognition
performance
across
the
conditions.
We
found

that
performance
in
the
No
Background
condition
was
significantly
lower
than

performance
in
the
Background
condition
(t(19)
=
2.520,
p
=
.021,
Cohen’s
d
=
1.107)
4
.

Thus,
performance
was
impaired
when
the
chicks
did
not
have
visual
experience
with
an

object
moving
along
patterned
backgrounds.

Analysis
of
Change
in
Performance
Over
Time

To
determine
whether
performance
changed
over
the
course
of
the
test
phase,
we

computed
a
repeated-‐measures
ANOVA
with
the
within-‐subjects
factor
of
Viewpoint

Change
and
Test
Day
and
the
between-‐subjects
factor
of
Condition
(see
Figure
10
for

performance
by
Test
Day).
The
ANOVA
revealed
a
significant
main
effect
of
Test
Day
(F(8,

4
The reported t-test includes test trials from the Background condition in which the objects were
presented on familiar backgrounds from the input phase. However, after removing the trials in
which the objects were presented on familiar backgrounds from the input phase, the t-test is still
significant.

47

152)
=
8.609,
p
<
10
-‐8
,
η
2

=
.312),
Viewpoint
Change
(Greenhouse-‐Geisser
adjusted,
F(1.533,

29.132)
=
5.936,
p
=
.011,
η
2

=
.238),
and
Condition
(F(1,
19)
=
6.590,
p
=
.019,
η
2

=
.258).

None
of
the
interactions
were
significant.

To
determine
whether
the
significant
main
effect
of
Test
Day
reflected
an

improvement
in
recognition
performance
over
time,
we
computed
a
post-‐hoc
correlation

between
test
day
and
performance,
which
revealed
a
significant
positive
correlation
for

both
conditions
(No
Background
condition:
r
=
.875,
p
=
.002;
Background
condition:
r
=

.814,
p
=
.008).
Thus,
chicks’
recognition
performance
improved
significantly
across
the
test

phase
as
they
acquired
more
experience
with
objects
moving
on
patterned
backgrounds.

Figure 9. Results. The graphs show the percent of time the chicks spent with the imprinted
object versus novel object for each condition and viewpoint change. The dashed line indicates
chance performance. Error bars show ±1 standard error (SEM). The chicks successfully
recognized their imprinted object across all background changes and viewpoint changes.
Overall, performance was lower in the No Background condition than the Background
condition.

48

Importantly,
chicks’
recognition
performance
was
still
well
above
chance
levels
even

on
the
first
test
day
in
the
Background
condition
(mean
=
68%,
SEM
=
3%,
two-‐tailed
one

sample
t-‐test:
t(9)
=
6.115,
p
<
.0002,
d
=
1.934).
However,
chicks’
recognition
performance

did
not
exceed
chance
levels
in
the
No
Background
condition
until
test
day
4
(all
Ps
Holm-‐
Bonferroni
corrected,
day
1:
t(10)
=
1.978,
corrected
p
=
.152;
day
2:
t(10)
=
1.811,

corrected
p
=
.100;
day
3:
t(10)
=
2.395,
corrected
p
=
.113;
day
4:
t(10)
=
3.421,
corrected
p

=
.026;
day
5:
t(10)
=
4.615,
corrected
p
=
.006;
day
6:
t(10)
=
4.269,
corrected
p
=
.008;
day

7:
t(10)
=
4.803,
corrected
p
=
.005;
day
8:
t(10)
=
4.903,
corrected
p
=
.005;
day
9:
t(10)
=

Figure 10. Overall performance and change over time. (A) The chicks’ overall object
recognition performance in each condition. Error bars show ±1 standard error (SEM). The
chicks in the Background condition performed significantly better than the chicks in No
Background condition. Critically, both groups of chicks received the same amount of
experience with object motion. (B) Average performance for each test day in the Background
condition (red line) and No Background condition (blue line). The dashed line indicates
chance performance. Shaded ribbons show ±1 standard error (SEM).

49

5.100,
corrected
p
=
.004).
Thus,
chicks
in
the
No
Background
condition
were
significantly

impaired
at
background-‐invariant
recognition
at
the
start
of
the
test
phase,
and
their

performance
improved
as
they
acquired
greater
amounts
of
experience
with
objects

moving
on
patterned
backgrounds
(i.e.,
during
the
test
trials).

Analysis
of
Individual
Subject
Performance

Since
we
collected
a
large
number
of
test
trials
from
each
subject,
we
were
able
to

analyze
each
chick’s
object
recognition
abilities
with
high
precision.
In
particular,
we

examined
whether
each
chick
was
able
to
build
a
background-‐invariant
representation
of

the
imprinted
object
(Figure
11).
In
the
Background
condition,
all
10
of
the
chicks

successfully
recognized
their
imprinted
object
across
the
test
phase
(two-‐tailed
one-‐sample

t-‐tests
with
Holm-‐Bonferroni
correction
for
multiple
comparisons,
all
corrected
Ps
<
.001)
5
,

indicating
that
all
of
the
chicks
built
a
background-‐invariant
object
representation.
In
the

No
Background
condition,
8
out
of
11
chicks
performed
significantly
above
chance
levels

(two-‐tailed
one-‐sample
t-‐tests
with
Holm-‐Bonferroni
correction
for
multiple
comparisons,

3
chicks:
p
>
.10;
remaining
8
chicks:
p
<
.05).

5
When the analysis is limited to test trials in which the objects are shown on novel backgrounds,
the results remain the same: all 10 chicks successfully recognized their imprinted object with
Holm-Bonferroni correction for multiple comparisons (all corrected Ps < .02).

50

Figure 11. Performance of the individual subjects. The graphs show each chick’s mean
recognition performance in the No Background condition (left, blue bars) and the Background
condition (right, red bars) across test days 1-4 (top) and test days 5-9 (bottom). The subjects
are ordered by average overall performance. Asterisks denote performance significantly above
chance levels after Holm-Bonferroni correction. The dashed line indicates chance
performance. Error bars show ±1 standard error (SEM). Most chicks in the No Background
Condition did not exceed chance performance in the first four days of testing but did succeed
in the last five days of testing. Nearly all chicks in the Background Condition exceeded
chance levels for the first four days of testing. In both experiments, there were significant
individual differences across subjects. Some newborn chicks developed better object
recognition abilities than others.

51

Discussion

We
used
an
automated
controlled-‐rearing
method
to
examine
whether
newborns

can
build
background-‐invariant
and
view-‐invariant
object
representations
at
the
onset
of

vision.
During
the
input
phase,
we
raised
newborn
chicks
in
strictly
controlled

environments
that
contained
a
single
virtual
object.
During
the
test
phase,
we
examined

whether
the
chicks
could
recognize
that
object
across
novel
backgrounds
and
novel

viewpoints.
Our
results
indicate
that
newborn
chicks
need
visual
experience
with
an
object

moving
across
patterned
backgrounds
in
order
to
develop
background-‐invariant

recognition.
When
newborn
chicks
were
raised
with
an
object
moving
on
patterned

backgrounds,
the
chicks
showed
robust
recognition
performance
across
the
entire
test

phase.
Conversely,
when
newborn
chicks
were
raised
with
an
object
moving
on
no

background,
the
chicks
were
impaired
at
recognizing
that
object
during
the
first
three
days

of
the
test
phase.
These
chicks
slowly
developed
background-‐invariant
object
recognition

as
they
acquired
greater
amounts
of
experience
with
objects
moving
across
patterned

backgrounds.
Moreover,
the
performance
of
the
chicks
in
both
conditions
improved

throughout
the
test
phase,
suggesting
that
background-‐invariant
object
recognition

continues
to
improve
across
the
first
two
weeks
of
life.

How
do
these
results
bear
on
the
classic
‘nature
versus
nurture’
debate?
The
present

study
demonstrates
that,
while
invariant
object
recognition
can
emerge
rapidly
in
the

newborn
brain,
specific
types
of
visual
experiences
are
necessary
to
develop
the
ability
to

segment
objects
from
backgrounds.
These
results
add
to
a
growing
body
of
work
showing

that
the
development
of
object
recognition
requires
specific
types
of
visual
experience
with

objects
(Wood,
2016;
Wood
et
al.,
2016;
Wood
&
Wood,
2016a).
In
particular,
newborn

52

chicks
need
visual
experience
with
objects
moving
slowly
and
smoothly
over
time
across

patterned
backgrounds
to
develop
object
recognition.
Thus,
newborn
chicks
can
build

abstract
object
representations,
but
only
when
provided
with
the
right
kind
of
visual
input.

The
present
study
also
informs
the
literature
on
the
development
of
visual
parsing.

Studies
of
infants
(Johnson
&
Aslin,
1995;
Johnson,
Bremner,
Slater,
Mason,
&
Foster,
2002;

Kellman
&
Spelke,
1983)
and
blind
patients
who
recover
sight
(Ostrovsky
et
al.,
2009)
have

demonstrated
that
motion
cues
are
critical
for
the
development
of
object
parsing
abilities.

The
present
study
extends
this
literature
by
demonstrating
that
motion
cues
alone
are
not

sufficient
for
newborn
visual
systems
to
segment
objects
from
backgrounds.
Rather,

newborns
need
visual
experience
with
objects
moving
on
backgrounds
to
build

background-‐invariant
object
representations.

To
conclude,
this
study
shows
that
newborn
chicks
can
segment
objects
from

backgrounds
within
the
first
few
days
of
life.
However,
the
development
of
this
ability

requires
visual
experience
with
objects
moving
on
patterned
backgrounds.
Thus,
this
study

reveals
a
constraint
on
the
development
of
object
recognition.
These
results
add
to
a

growing
body
of
work
showing
that
the
development
of
object
recognition
requires
visual

experience
with
a
natural
visual
environment,
containing
objects
that
move
slowly
and

smoothly
over
time
over
patterned
visual
scenes.
In
the
absence
of
such
natural
visual

experience,
newborn
animals
develop
impaired
object
recognition
abilities.

53

Chapter
4:
Newborn
chicks
generate
view-‐invariant
object
representations
from
sparse

visual
input

(Corresponding
publication:
Wood,
S.
M.
W.
&
Wood,
J.
N.
(2015)
A
chicken
model
for

studying
the
emergence
of
invariant
object
recognition.
Frontiers
in
Neural
Circuits,
9,
7.)

Abstract

“Invariant
object
recognition”
refers
to
the
ability
to
recognize
objects
across
variation
in

their
appearance
on
the
retina.
This
ability
is
central
to
visual
perception,
yet
its

developmental
origins
are
poorly
understood.
Traditionally,
nonhuman
primates,
rats,
and

pigeons
have
been
the
most
commonly
used
animal
models
for
studying
invariant
object

recognition.
Although
these
animals
have
many
advantages
as
model
systems,
they
are
not

well
suited
for
studying
the
emergence
of
invariant
object
recognition
in
the
newborn

brain.
Here,
we
argue
that
newborn
chicks
(Gallus
gallus)
are
an
ideal
model
system
for

studying
the
emergence
of
invariant
object
recognition.
Using
an
automated
controlled-‐
rearing
approach,
we
show
that
chicks
can
build
a
view-‐invariant
representation
of
the
first

object
they
see
in
their
life.
This
invariant
representation
can
be
built
from
highly

impoverished
visual
input
(3
images
of
an
object
separated
by
15°
azimuth
rotations)
and

cannot
be
accounted
for
by
low-‐level
retina-‐like
or
V1-‐like
neuronal
representations.
These

results
indicate
that
newborn
neural
circuits
begin
building
invariant
object

representations
at
the
onset
of
vision
and
argue
for
an
increased
focus
on
chicks
as
animal

models
for
studying
invariant
object
recognition.

54

Introduction

Humans
and
other
animals
can
recognize
objects
despite
tremendous
variation
in

how
objects
appear
on
the
retina
(due
to
changes
in
viewpoint,
size,
lighting,
and
so
forth).

This
ability—known
as
“invariant
object
recognition”
6
—has
been
studied
extensively
in

adult
animals,
but
its
developmental
origins
are
poorly
understood.
We
have
not
yet

characterized
the
initial
state
of
object
recognition
(i.e.,
the
state
of
object
recognition
at

the
onset
of
vision),
nor
do
we
understand
how
this
initial
state
changes
as
a
function
of

specific
visual
experiences.

Researchers
have
long
recognized
that
studies
of
newborns
are
essential
for

characterizing
the
initial
state
of
visual
cognition;
however,
methodological
constraints

have
hindered
our
ability
to
study
invariant
object
recognition
in
newborn
humans.
First,

human
infants
cannot
ethically
be
raised
in
controlled
environments
from
birth.

Consequently,
researchers
have
been
unable
to
study
how
specific
visual
experiences
shape

the
initial
state
of
invariant
object
recognition.
Second,
it
is
typically
possible
to
collect
just

a
small
number
of
test
trials
from
each
newborn
human.
As
a
result,
researchers
have
been

unable
to
measure
newborns’
first
visual
object
representations
with
high
precision.
The

recent
development
of
an
automated
controlled-‐rearing
approach
with
a
newborn
7

animal

model—the
domestic
chick
(Gallus
gallus)—overcomes
these
two
limitations.

Recent
studies
using
this
automated
controlled-‐rearing
method
have
revealed
that

newborn
chicks
are
capable
of
building
invariant
representations
of
objects
and
faces
from

6
Here, we use “invariant” to mean tolerant to changes in appearance, but not necessarily fully
invariant (i.e., recognizable across any viewing condition or without performance costs for
changes in viewing condition).
7
The term “newborn” is used to refer to an animal at the beginning of the post-embryonic phase
of their life cycle.

55

sparse
visual
experience.
For
example,
newborn
chicks
raised
with
a
single
object
(Wood,

2013)
or
face
(Wood
&
Wood,
in
prep)
rotating
through
a
single
viewpoint
range
can
build

a
view-‐invariant
representation
of
that
object/face.
Similarly,
newborn
chicks
that
are

raised
with
a
single
object
rotating
on
a
single
background
can
build
a
background-‐
invariant
representation
of
that
object
(Chapter
2).
While
newborn
chicks
can
build

representations
that
generalize
well
beyond
their
visual
experiences,
there
are
limits
to
the

development
of
this
ability.
To
illustrate,
when
newborn
chicks
were
raised
without
visual

experience
of
an
object
moving
along
patterned
backgrounds
(i.e.,
an
object
that
rotated
on

a
white
homogenous
background),
the
chicks
were
significantly
impaired
at
background-‐
invariant
recognition
and
failed
to
recognize
the
imprinted
object
above
chance
levels
on

the
first
days
of
testing
(Chapter
3).
Thus,
it
is
possible
that
visual
object
experiences
that

are
too
sparse
can
impair
the
development
of
object
recognition.
Overall,
however,
the

limits
on
newborns’
ability
to
generalize
from
sparse
visual
experiences
remain
unclear.

The
Present
Experiment

The
current
study
builds
on
a
previous
study
that
examined
whether
newborn

chicks
can
build
invariant
object
representations
at
the
onset
of
vision
(Wood,
2013).
In

this
previous
study,
chicks
were
raised
for
one
week
in
environments
that
contained
a

single
virtual
object
that
could
only
be
seen
from
a
limited
60°
viewpoint
range.
In
their

second
week
of
life,
Wood
(2013)
then
measured
whether
chicks
could
recognize
the

virtual
object
across
a
variety
of
novel
viewpoints.
The
majority
of
subjects
successfully

recognized
the
object
across
the
novel
viewpoints,
which
shows
that
chicks
can
build
a

view-‐invariant
representation
of
the
first
object
they
see
in
their
life.

56

The
present
study
extends
this
finding
in
three
ways.
First,
we
significantly
reduced

the
amount
of
visual
object
input
available
to
the
subjects.
In
Wood
(2013),
the
chicks
were

shown
a
virtual
object
that
moved
smoothly
over
time
through
a
60°
viewpoint
range
at
24

images/second,
whereas
in
the
present
study,
the
chicks
were
shown
a
virtual
object
that

moved
abruptly
over
time
through
a
30°
viewpoint
range
at
1
image/second
(see
Figure

12).
Thus,
compared
with
Wood
(2013),
the
chicks
in
the
present
study
observed
a
smaller

number
of
unique
images
of
the
object
(3
unique
images
versus
72
unique
images),
a

smaller
range
of
movement
(30°
viewpoint
range
versus
60°
viewpoint
range),
and

unnatural
(abrupt)
versus
natural
(smooth)
object
motion.
The
abrupt
object
motion
was

unnatural
because
it
caused
the
object’s
features
to
move
large
distances
across
the
retina

instantaneously,
breaking
the
spatiotemporal
contiguity
of
the
images.
The
present
study

therefore
provided
a
particularly
strong
test
of
whether
chicks
can
build
invariant
object

representations
from
impoverished
visual
input.

Second,
we
tested
chicks’
object
recognition
abilities
across
a
systematically
varying

recognition
space.
Each
chick’s
object
recognition
abilities
were
tested
across
27
different

viewpoint
ranges;
the
viewpoint
ranges
canvassed
a
uniform
recognition
space
in
which

the
object
was
rotated
-‐60°
to
+60°
in
the
azimuth
direction
and
-‐60°
to
+60°
in
the

elevation
direction
(in
15°
increments;
see
Figure
14).
Thus,
we
were
able
to
examine

whether
chicks’
recognition
performance
varied
as
a
function
of
the
object’s
degree
of

rotation.

Third,
we
investigated
whether
chicks’
recognition
abilities
could
be
explained
by

some
low-‐level
features
of
the
test
animations,
by
quantifying
the
similarity
between
the

input
images
and
the
test
images.
We
quantified
image
similarity
in
terms
of
both
pixel-‐like

57

similarity
and
V1-‐like
similarity,
akin
to
previous
studies
that
tested
object
recognition
in

adult
rats
(Tafazoli
et
al.,
2012;
Zoccolan
et
al.,
2009).

Methods

Subjects

Ten
chicks
of
unknown
sex
were
tested.
No
subjects
were
excluded
from
the

analyses.
Fertilized
eggs
were
incubated
in
darkness
in
an
OVA-‐Easy
incubator
(Brinsea

Products
Inc.,
Titusville,
FL).
We
maintained
the
temperature
and
humidity
at
99.6°F
and

45%,
respectively,
for
the
first
19
days
of
incubation.
On
day
19
of
incubation,
the
humidity

was
increased
to
60%.
The
eggs
were
incubated
in
darkness
to
ensure
that
no
visual
input

would
reach
the
chicks
through
their
shells.
After
hatching,
we
moved
the
chicks
from
the

incubator
room
to
the
controlled-‐rearing
chambers
in
complete
darkness.
Each
chick
was

raised
singly
within
its
own
chamber.

Controlled-‐Rearing
Chambers

The
controlled-‐rearing
chambers
measured
66
cm
(length)
×
42
cm
(width)
×
69
cm

(height).
The
floors
of
the
chambers
consisted
of
black
wire
mesh
suspended
1”
over
a

black
surface
by
transparent,
plexiglass
beams.
Object
stimuli
were
presented
to
the

subjects
by
projecting
virtual
objects
onto
two
display
walls
(19”
LCD
monitors
with
1440

×
900
pixel
resolution)
situated
on
opposite
sides
of
the
chambers.
The
other
two
walls
of

the
chambers
were
white,
high-‐density
plastic.
We
used
matte
(non-‐reflective)
materials

for
both
the
walls
and
the
floor
to
avoid
incidental
illumination.
The
chambers
contained

58

no
rigid,
bounded
objects
other
than
the
virtual
objects
presented
on
the
display
walls.
See

Figure
1
in
Wood
(2013)
for
a
picture
of
the
chambers.

Food
and
water
were
provided
ad
libitum
within
transparent,
rectangular
troughs
in

the
ground
(66
cm
length
×
2.5
cm
width
×
2.7
cm
height).
Grain
was
used
as
food
because

grain
does
not
behave
like
a
rigid,
bounded
object
(i.e.,
grain
does
not
maintain
a
solid,

bounded
shape).
All
care
of
the
chicks
was
performed
in
darkness
with
the
aid
of
night

vision
goggles.

The
controlled-‐rearing
chambers
recorded
all
of
the
chicks’
behavior
(24
hours/day,

7
days/week)
with
high
precision
(9
samples/second)
via
micro-‐cameras
(1.5
cm

diameter)
embedded
in
the
ceilings
of
the
chambers
and
automated
image-‐based
tracking

software
(Ethovision
XT,
Noldus
Information
Technology,
Leesburg,
VA).
This
software

calculated
the
amount
of
time
each
chick
spent
within
zones
(22
cm
×
42
cm)
next
to
each

display
wall.
In
total,
3,360
hours
of
video
footage
(14
days
×
24
hours/day
×
10
subjects)

were
collected
and
analyzed
for
the
present
study.

Input
Phase

During
the
input
phase
(the
first
week
of
life),
chicks
were
raised
in
environments

that
contained
a
single
virtual
object.
Four
chicks
were
presented
with
Object
1
and
six

chicks
were
presented
with
Object
2
(see
Figure
12).
The
object
animations
contained
just

three
unique
images
of
the
object:
a
front
view
and
two
side
views
with
±15°
azimuth

rotations.
The
images
changed
at
a
rate
of
1
image/sec.
From
a
human
adult’s
perspective,

the
objects
appeared
to
undergo
apparent
motion,
rocking
back
and
forth
through
a
30°

viewpoint
range
along
a
frontoparallel
vertical
axis.
The
virtual
object
was
displayed
on
a

59

uniform
white
background,
and
appeared
for
an
equal
amount
of
time
on
the
left
and
right

display
walls.
The
object
switched
walls
every
two
hours,
following
a
one-‐minute
period
of

darkness
(Figure
13).

Test
Phase

During
the
test
phase
(the
second
week
of
life),
we
examined
whether
each
chick

had
built
a
view-‐invariant
representation
of
their
imprinted
object
by
using
an
automated

two-‐alternative
forced
choice
testing
procedure.
On
each
test
trial,
the
imprinted
object

was
shown
on
one
display
wall
and
an
unfamiliar
object
was
shown
on
the
other
display

Figure 12. The three unique images of Object 1 and Object 2 presented to the chicks
during the input phase. Four chicks were presented with Object 1 and six chicks were
presented with Object 2. Object 2 served as the unfamiliar object for the chicks that
were imprinted to Object 1, and vice versa. The three images changed at a rate of 1
image/second, causing the virtual object to rotate abruptly back and forth through a 30°
viewpoint range. Chicks never observed the virtual object (or any other object) from
another viewpoint during the input phase.

60

wall.
We
then
measured
the
amount
of
time
chicks
spent
in
proximity
to
each
object.
If

chicks
successfully
recognize
their
imprinted
object,
then
they
should
spend
a
greater

proportion
of
time
in
proximity
to
the
imprinted
object
compared
to
the
unfamiliar
object.

The
imprinted
object
was
shown
from
81
different
test
viewpoints,
consisting
of
all

possible
combinations
of
9
azimuth
rotations
(-‐60°,
-‐45°,
-‐30°,
-‐15°,
0°,
+15°,
+30°,
+45°,

+60°)
and
9
elevation
rotations
(-‐60°,
-‐45°,
-‐30°,
-‐15°,
0°,
+15°,
+30°,
+45°,
+60°).
To
equate

the
direction
of
object
motion
across
the
input
and
test
phases,
the
81
viewpoints
were

organized
into
27
different
viewpoint
ranges,
each
containing
three
images.
Like
the
input

Figure 13. A schematic showing how the virtual objects were presented on the two display
walls during the input phase (top) and the test phase (bottom). During the input phase, chicks
observed a single virtual object rotating abruptly back and forth through a 30° viewpoint
range. During the test phase, chicks were presented with regularly scheduled test trials.
During the test trials, the imprinted object was shown on one display wall and an unfamiliar
object was shown on the other display wall. The imprinted object was shown from a variety of
novel viewpoints, whereas the unfamiliar object was always shown from the same frontal
viewpoint range as the imprinted object during the input phase. This maximized the pixel-
level and V1-level similarity between the unfamiliar object and the imprinting stimulus. Thus,
to recognize their imprinted object, chicks needed to generalize across large, novel, and
complex changes in the object’s appearance on the retina.

61

object
animation,
each
of
the
27
test
animations
showed
the
imprinted
object
rotating
back

and
forth
±15°
along
the
azimuth
rotation
axis.
Figure
15
shows
how
the
81
individual

viewpoints
were
organized
into
the
27
test
animations.

The
unfamiliar
object
was
similar

to
the
imprinted
object
in
terms
of
its
size,
color,
motion
speed,
and
motion
trajectory.

Further,
on
all
of
the
test
trials,
the
unfamiliar
object
was
presented
from
the
same
frontal

viewpoint
range
as
the
imprinted
object
from
the
input
phase.
Presenting
the
unfamiliar

object
from
this
frontal
viewpoint
range
maximized
the
similarity
between
the
unfamiliar

object
and
the
imprinting
stimulus.
Thus,
to
recognize
their
imprinted
object,
chicks

needed
to
generalize
across
large,
novel,
and
complex
changes
in
the
object’s
appearance

on
the
retina.
The
test
trials
lasted
17
minutes
and
were
separated
from
one
another
by
32-‐
minute
rest
periods.
During
the
rest
periods,
we
projected
the
animation
from
the
input

phase
onto
one
display
wall
and
a
white
screen
onto
the
other
display
wall.
The
test
trials

and
rest
periods
were
separated
by
1-‐minute
periods
of
darkness.
On
each
day
of
the
test

phase,
chicks
were
presented
with
each
viewpoint
range
one
time,
for
a
total
of
27
test

trials
per
day.
Thus,
each
chick
received
189
test
trials
over
the
course
of
the
experiment.

The
27
viewpoint
ranges
were
presented
in
a
randomized
order
during
each
day
of
the
test

phase.

Results

Overall
Performance

To
test
whether
performance
was
significantly
above
chance,
we
used
intercept-‐
only
mixed
effects
models
(also
called
“multilevel
models”).
Since
we
collected
multiple

observations
from
each
subject,
it
was
necessary
to
use
an
analysis
that
can
account
for
the

62

nested
structure
of
the
data
(Aarts,
Verhage,
Veenvliet,
Dolan,
&
van
der
Sluis,
2014).
The

mixed
effects
models
were
performed
using
R
(www.r-‐project.org).
First,
we
computed
the

number
of
test
trials
in
which
chicks
preferred
their
imprinted
object
over
the
unfamiliar

object.
The
chick
was
rated
to
have
preferred
their
imprinted
object
on
a
trial
if
their
object

preference
score
was
greater
than
50%.
The
object
preference
score
was
calculated
with

the
formula:

Accordingly,
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.
Chicks
spent
more
time
with
their
imprinted
object
on

59%
(SEM
=
3%)
of
the
test
trials
(see
Figure
14).

We
used
a
mixed
effects
logistic
regression
model
(R
package
lme4)
to
test
whether

performance
was
significantly
greater
than
chance.
We
fitted
the
model
with
test
trial

outcome
(binary:
correct
or
incorrect)
as
the
dependent
variable,
an
intercept
as
the
fixed

effect,
and
a
random
intercept
for
the
subject-‐effect.
The
fixed
effect
intercept
was
positive

and
significant
(b
=
0.394,
z
=
2.857,
p
=
0.004),
which
indicates
that
chicks’
recognition

performance
was
significantly
greater
than
50%
(chance
performance).
Chicks’
recognition

performance
was
also
significantly
above
chance
when
the
analysis
did
not
include
the
test

trials
where
the
imprinted
object
was
shown
from
the
familiar
viewpoint
range
(b
=
0.365,

z
=
2.637,
p
=
0.008).

Object
Preference
Score

=

Time
by
Imprinted
Object

Time
by
Imprinted
Object
+
Time
by
Unfamiliar
Object

63

Second,
we
confirmed
these
results
with
a
similar
analysis
on
the
object
preference

scores
(i.e.,
the
proportion
of
time
chicks
spent
with
the
imprinted
object
compared
to
the

unfamiliar
object).
Because
the
significance
of
the
intercept
indicates
whether
the
intercept

is
significantly
different
than
0,
we
subtracted
50%
from
each
object
preference
score.

Thus,
the
adjusted
object
preference
scores
ranged
from
-‐50%
to
+50%,
with
an
adjusted

Figure
14.
Recognition
performance
for
the
overall
group
(top)
and
the
individual

subjects
(bottom).
The
dark
gray
bars
denote
the
percentage
of
correct
trials,
and
the
light

gray
bars
denote
the
proportion
of
time
subjects
spent
with
the
imprinted
object.
These

graphs
do
not
include
the
test
trials
in
which
the
imprinted
object
was
shown
from
the

familiar
viewpoint
range
from
the
input
phase.
The
subjects
are
ordered
by
performance.

The
red
dashed
lines
show
chance
performance
(50%).
P-‐values
denote
the
statistical

difference
between
the
number
of
correct
and
incorrect
test
trials
as
computed
through

one-‐tailed
binomial
tests.

64

object
preference
score
of
0
indicating
equal
time
spent
with
the
imprinted
object
and

unfamiliar
object.
We
fitted
a
linear
mixed
effects
model
(R
package
nlme)
with
the

adjusted
object
preference
score
as
the
dependent
variable,
an
intercept
as
the
fixed
effect,

and
a
random
intercept
for
the
subject-‐effect.
Again,
the
fixed
effect
intercept
was
positive

and
significant
(b
=
0.072,
t(1878)
=
3.015,
p
=
0.003),
which
provides
further
evidence
that

chicks’
recognition
performance
was
significantly
higher
than
50%
(chance
performance).

Chicks’
recognition
performance
was
also
significantly
above
chance
when
the
analysis
did

not
include
the
test
trials
where
the
imprinted
object
was
shown
from
the
familiar

viewpoint
range
(b
=
0.068,
t(1808)
=
2.828,
p
=
0.005).

With
this
controlled-‐rearing
method
we
were
able
to
collect
a
large
number
of
test

trials
from
each
chick.
Thus,
we
were
able
to
examine
whether
each
subject
was
able
to

build
a
view-‐invariant
representation
of
their
imprinted
object.
To
do
so,
we
computed

whether
each
subject’s
performance
across
the
test
trials
exceeded
chance
level
(using
one-‐
tailed
binomial
tests).
Six
of
the
10
subjects
successfully
built
an
invariant
object

representation
(Ps
≤
0.05).
8

When
the
analysis
did
not
include
the
familiar
viewpoint
range

from
the
input
phase,
5
of
the
10
chicks
performed
significantly
above
chance
(see
Figure

14).
Thus,
many
of
the
chicks
successfully
built
an
invariant
object
representation
that

generalized
across
novel
viewpoints.

8
Four of the 10 subjects performed significantly higher than chance level after a Bonferroni
correction for 10 independent tests (10 subjects; p < 0.005).

65

To
ensure
that
all
of
the
chicks
successfully
imprinted
to
the
virtual
object
(i.e.,

developed
an
attachment
to
the
object),
we
examined
whether
the
chicks
showed
a

preference
for
the
imprinted
object
during
the
rest
periods
in
the
test
phase.
All
10
subjects

Figure 15. (Top) The test viewpoints shown during the test phase. The viewpoint range shown
during the input phase is indicated by the blue frames. (Bottom) Chicks’ average percentage of
correct trials across the 27 viewpoint ranges. Chance performance was 50%. Each subject
received 7 test trials for each viewpoint range. Thus, each viewpoint cell in the figure reflects
the data from 28 test trials for Object 1 (7 test trials × 4 subjects) and 42 test trials for Object 2
(7 test trials × 6 subjects), for a total of 1,890 test trials across all viewpoint ranges.

66

spent
the
majority
of
the
rest
periods
in
proximity
to
the
imprinting
stimulus
(mean
=
88%

of
trials;
SEM
=
2%;
one-‐tailed
binomial
tests,
all
P
<
10
−9
).
Thus,
it
is
possible
to
imprint
to

an
object
but
fail
to
build
a
view-‐invariant
representation
of
that
object
(see
also
Wood,

2013).

Correlations
of
Object
Recognition
Performance
Across
Subjects

As
shown
in
Figure
14,
there
was
substantial
variation
in
chicks’
recognition

abilities.
To
examine
whether
chicks’
recognition
abilities
were
correlated
with
one

another,
we
measured
the
correlation
in
performance
across
the
viewpoint
ranges
for
each

pair
of
chicks.
Specifically,
we
computed
the
percentage
of
time
spent
with
the
imprinted

object
for
each
viewpoint
range
for
each
chick.
The
correlations
in
performance
between
all

pairs
of
chicks
are
shown
in
Figure
16.
Performance
was
highly
correlated
across
the

subjects:
out
of
the
45
subject
pairs,
44
were
positively
correlated
and
only
1
pair
was

negatively
correlated.
Overall,
the
average
correlation
between
subjects
was
r
=
0.35
(SEM

=
0.03).
These
correlation
values
were
significantly
different
from
0
(no
correlation),
t(44)

=
8.72,
p
<
0.001.
Despite
the
substantial
range
of
variation
in
performance
across
subjects,

the
chicks’
recognition
abilities
were
nevertheless
highly
correlated
with
one
another.

Analysis
of
Change
in
Performance
Over
Time

To
examine
whether
recognition
performance
changed
over
the
course
of
the
test

phase,
we
calculated
the
percentage
of
time
chicks
spent
in
proximity
to
the
imprinted

object
versus
the
unfamiliar
object
as
a
function
of
test
day.
The
results
are
shown
in
Figure

17.
Performance
remained
stable
across
the
test
phase
(one-‐way
ANOVA,
F(6)
=
0.224,
p
=

67

0.968).
Chicks’
recognition
behavior
was
spontaneous
and
robust,
and
cannot
be
explained

by
learning
taking
place
across
the
test
phase.
Chicks
immediately
achieved
their
maximal

performance
and
did
not
significantly
improve
thereafter.

Analysis
of
Viewpoint
Effects

To
test
whether
recognition
performance
varied
as
a
function
of
the
degree
of

viewpoint
change,
we
calculated
chicks’
mean
object
preference
scores
for
each
of
the

elevation
viewpoint
change
magnitudes
(i.e.,
±60°,
±45°,
±30°,
±15°,
0°).
The
correlation

between
the
magnitude
of
viewpoint
change
and
performance
did
not
approach

Figure 16. A similarity matrix showing the correlation in performance for each pair of subjects.
The order of the subjects in the matrix is determined by a hierarchical cluster analysis. The cells
are color-coded by correlation value: green values = positive correlation in performance; red
values = negative correlation in performance. The color scale reflects the full range of possible
correlation values.

68

significance
(r
=
-‐0.06,
p
=
0.93).
Thus,
when
chicks
first
begin
to
recognize
objects,
their

performance
does
not
decline
with
larger
changes
in
viewpoint.

Analysis
of
Object
Stimuli
and
Performance

Did
chicks
need
high-‐level
(invariant)
object
representations
to
succeed
in
this

experiment?
Previous
studies
have
shown
that
chicks
do
not
use
overall
brightness
as
a

low-‐level
cue
to
distinguish
between
these
two
virtual
objects
(Wood,
2014a),
and
that

chicks’
early
emerging
invariant
object
recognition
abilities
cannot
be
explained
by
retina-‐
like
(pixel-‐wise)
representations
when
recognition
is
tested
across
more
extreme
azimuth

and
elevation
rotations
(Wood,
2013).

Figure 17. Change in chicks’ object recognition performance over time. The graph illustrates
group mean performance over the full set of viewpoint ranges shown during the 7-day test
phase, computed for the first, second, third, etc., day of testing. Chance performance was 50%.
Chicks’ recognition performance did not change significantly across the course of the test
phase.

69

To
extend
these
previous
analyses,
we
quantified
the
similarity
between
the
input

animations
and
the
test
animations
in
two
ways.
First,
we
computed
the
amount
of
image

variation
between
the
input
animations
and
the
test
animations
from
a
retina-‐like
(pixel-‐
level)
perspective.
For
each
animation,
we
(1)
measured
the
brightness
level
of
each
pixel

in
each
of
the
3
unique
object
images,
(2)
compared
each
image
from
the
test
animation
to

each
image
from
the
input
animation
(i.e.,
by
comparing
the
brightness
level
of
each

corresponding
pixel
across
the
images
and
taking
the
absolute
difference),
and
(3)

calculated
the
average
pixel-‐level
difference
between
the
three
unique
images
from
the

input
and
test
animations
(i.e.,
the
1st
test
image
was
compared
to
the
1st,
2nd,
and
3rd

input
image;
the
2nd
test
image
was
compared
to
the
1st,
2nd,
and
3rd
input
image;
and
the

3rd
test
image
was
compared
to
the
1st,
2nd,
and
3rd
input
image).
Recognition

performance
(i.e.,
the
object
preference
scores)
did
not
vary
as
a
function
of
the
pixel-‐level

difference
between
the
input
animations
and
test
animations
(linear
regression:
b
=
-‐
7.08×10
-‐8
,
t(52)
=
-‐1.29,
p
=
0.20).

Second,
we
computed
the
amount
of
image
variation
between
the
input
animations

and
the
test
animations
from
a
V1-‐level
perspective.
To
do
so,
we
used
a
Gabor
measure
of

similarity
with
the
Gabor
jet
model:
a
multi-‐scale,
multi-‐orientation
model
of
V1
complex-‐
cell
filtering
developed
by
Lades
et
al.
(1993).
The
general
parameters
and
implementation

followed
those
used
by
Xu
&
Biederman
(2010),
which
can
be
downloaded
at

http://geon.usc.edu/GWTgrid_simple.m.
For
each
unique
image
in
each
animation,
we

measured
the
magnitude
of
activation
values
that
the
image
produced
in
a
set
of
40
Gabor

jets
(8
orientations
×
5
scales).
We
measured
the
dissimilarity
between
two
images
by

computing
1
minus
the
correlation
between
their
Gabor
jet
activation
values.
Thus,
the

70

dissimilarity
between
two
images
could
range
from
0
(perfect
positive
correlation)
to
2

(perfect
negative
correlation).
Finally,
we
calculated
the
average
Gabor
jet
dissimilarity

across
all
three
unique
images
of
the
animations
(i.e.,
the
1st
test
image
was
compared
to

the
1st,
2nd,
and
3rd
input
image;
the
2nd
test
image
was
compared
to
the
1st,
2nd,
and
3rd

input
image;
and
the
3rd
test
image
was
compared
to
the
1st,
2nd,
and
3rd
input
image).

Recognition
performance
(i.e.,
the
object
preference
scores)
did
not
vary
as
a
function
of

Gabor
jet
dissimilarity
between
the
input
animations
and
test
animations
(linear

regression:
b
=
-‐0.11,
t(52)
=
-‐1.04,
p

=
0.30).

Additionally,
to
confirm
that
chicks’
recognition
performance
could
not
be
explained

by
retina–like
or
V1–like
representations,
we
tested
whether
models
based
on
pixel-‐level

or
V1-‐level
representations
could
successfully
predict
object
identity
in
this
experiment.

Specifically,
we
generated
a
pixel-‐level
model
and
a
V1-‐level
model
that
predicted
object

identity
based
on
the
image
differences
between
the
test
animations
and
the
input

animation.
For
each
viewpoint
range,
we
measured
(1)
the
difference
between
the
test

animation
of
the
imprinted
object
and
the
input
animation
of
the
imprinted
object
(within-‐
object
difference),
and
(2)
the
difference
between
the
test
animation
of
the
unfamiliar

object
and
the
input
animation
of
the
imprinted
object
(between-‐object
difference;
see

Figure
18).
If
the
within-‐object
difference
was
smaller
than
the
between-‐object
difference,

then
the
model
was
“correct”
for
that
viewpoint
range.
Conversely,
if
the
between-‐object

difference
was
smaller
than
the
within-‐object
difference,
then
the
model
was
“incorrect”
for

that
viewpoint
range.
The
retina-‐like
(pixel-‐level)
model
was
correct
for
20%
of
the

viewpoint
ranges,
while
the
V1-‐level
(Gabor
jet)
model
was
correct
for
28%
of
the

viewpoint
ranges.
Unlike
the
chicks’
recognition
performance,
which
was
significantly

71

above
chance
(50%)
levels,
both
low-‐level
models
performed
significantly
below
chance

levels
(pixel-‐level
intercept-‐only
logistic
regression:
b
=
-‐1.36,
z
=
-‐4.04,
p
<
0.0001;
V1-‐
level
intercept-‐only
logistic
regression:
b
=
-‐0.96,
z
=
-‐3.15,
p
=
0.002).

Figure 18. The average pixel-level and V1-level differences between the three unique images
of each test animation and the three unique images of the input animation (i.e., the 1st test
image was compared to the 1st, 2nd, and 3rd input image; the 2nd test image was compared to
the 1st, 2nd, and 3rd input image; and the 3rd test image was compared to the 1st, 2nd, and
3rd input image). The orange bars show the between-object differences (i.e., the difference
between the test animation of the unfamiliar object and the input animation of the imprinted
object). The blue bars (ordered by similarity) show the within-object differences (i.e., the
difference between the test animation of the imprinted object and the input animation of the
imprinted object). The top graphs show the differences as measured at the pixel-level, and the
bottom graphs show the differences as measured at the V1-level (using Gabor jet
magnitudes). Overall, the within-object difference was less than the between-object difference
on only 20% (pixel-level) and 28% (V1-level) of the viewpoint ranges (chance performance =
50%). Thus, neither pixel-level nor V1-level representations can be used to reliably predict
object identity in this experiment.

72

To
compare
the
models’
performance
to
the
chicks’
performance,
we
computed
the

average
percentage
of
time
chicks
spent
with
the
imprinted
object
versus
the
unfamiliar

object
for
each
viewpoint
range.
If
chicks
spent
more
time,
on
average,
with
the
imprinted

object
than
the
unfamiliar
object,
then
the
chicks
were
“correct”
for
that
viewpoint
range.

Conversely,
if
chicks
spent
more
time
with
the
unfamiliar
object
than
the
imprinted
object,

then
the
chicks
were
“incorrect”
for
that
viewpoint
range.
For
each
model
and
for
the

chicks,
there
were
54
conditions
(27
viewpoint
ranges
×
2
imprinted
objects).
The
chicks

were
correct
on
35
conditions
and
incorrect
on
19
conditions.
The
pixel-‐level
model
was

correct
on
11
conditions
and
incorrect
on
43
conditions.
The
V1-‐level
model
was
correct
on

15
conditions
and
incorrect
on
39
conditions.
Chi-‐square
tests
comparing
the
number
of

correct
and
incorrect
conditions
for
the
chicks
and
the
models
found
significant
differences

between
chicks’
recognition
performance
and
both
models’
recognition
performance

(pixel-‐level
model
versus
chick
performance:
X
2
(1,
N
=
108)
=
21.81,
p
<
10
-‐5
;
V1-‐level

model
versus
chick
performance:
X
2
(1,
N
=
108)
=
14.90,
p
<
10
-‐3
).

Overall,
the
within-‐object
difference
was
greater
than
the
between-‐object
difference,

both
at
the
pixel-‐level
and
V1-‐levels.
Thus,
in
principle,
chicks
could
have
succeeded
in
this

experiment
by
preferring
the
test
animation
that
was
the
most
different
from
the
input

animation
(i.e.,
a
novelty
preference).
To
test
this
possibility,
we
analyzed
the
test
trials
in

which
the
imprinted
object
was
presented
from
the
familiar
viewpoint
range
from
the

input
phase.
If
chicks
had
a
novelty
preference,
then
they
should
have
avoided
the

imprinted
object
on
the
trials
in
which
the
test
animation
of
the
imprinted
object
was

identical
to
the
input
animation
of
the
imprinted
object.
Contrary
to
this
prediction,
chicks

spent
significantly
more
time
with
the
imprinted
object
than
the
unfamiliar
object
when

73

the
imprinted
object
was
presented
from
the
familiar
viewpoint
range
(logistic
mixed

effects
regression:
b
=
1.514,
z
=
3.229,
p
=
0.001;
linear
mixed
effects
regression:
b
=
0.180,

t(60)
=
3.062,
p
=
0.003).
Thus,
chicks
did
not
simply
have
a
preference
for
the
novel

animation
in
this
experiment.

Together,
these
analyses
indicate
that
chicks
build
invariant
object
representations

that
cannot
be
explained
by
low-‐level
retina-‐like
(pixel-‐wise)
or
V1-‐like
neuronal

representations.
Rather,
chicks
build
selective
and
tolerant
object
representations,
akin
to

those
found
in
higher
levels
of
the
visual
system.

Effects
of
self-‐occlusion

If
the
chicks
were
not
relying
on
low-‐level
retina-‐like
or
V1-‐like
representations,

what
types
of
representations
did
they
form?
What
explains
the
variation
in
chicks’

performance
across
different
viewpoints?
One
possibility
is
that
the
chicks
formed

representations
of
the
parts
comprising
their
imprinted
object.
Thus,
when
an
object
is

highly
self-‐occluded,
fewer
parts
of
the
object
are
visible,
and
recognition
should
be
more

difficult.
Under
conditions
of
self-‐occlusion,
discriminative
features
that
could
be
used
to

recognize
an
object
may
not
be
visible.
Consistent
with
this
theory,
chicks’
recognition

performance
was
generally
lower
when
the
object
was
presented
from
negative
elevation

rotations
(see
Figure
15).
When
the
object
was
presented
from
negative
elevation

rotations,
a
smaller
portion
of
the
object
was
visible
to
the
subject
(see
Figure
15).
To

provide
a
quantitative
measurement
of
self-‐occlusion
in
the
test
animations,
we
computed

the
number
of
foreground
(object)
pixels
that
were
visible
on
the
screen
for
each

animation.
We
found
that
chicks’
recognition
performance
(i.e.,
the
percentage
of
time

74

spent
with
the
imprinted
object
versus
unfamiliar
object)
was
positively
correlated
with

the
number
of
foreground
(object)
pixels
that
were
visible
on
the
screen
(r
=
0.41,
p
<
0.01).

This
result
is
consistent
with
a
recent
study
of
adult
rats
who
were
trained
to

distinguish
between
these
same
two
virtual
objects
(Alemi-‐Neissi
et
al.,
2013).
Alemi-‐Neissi

et
al.
found
that
rats
built
sub-‐features
of
objects
that
were
smaller
than
the
entire
object.

When
these
sub-‐features
were
occluded
with
“bubble
masks”
(Gosselin
&
Schyns,
2001),

rats’
recognition
abilities
declined.
It
would
be
interesting
for
future
studies
to
use
this

bubble
masking
approach
with
chicks
to
characterize
the
specific
features
used
to

recognize
objects
at
the
onset
of
vision.

Comparison
to
Prior
Studies

The
virtual
objects
used
in
this
study
were
the
same
as
those
used
in
Wood
(2013).

However,
in
the
current
study,
each
imprinting
and
test
animation
only
contained
3
unique

images
showing
the
objects
rotating
abruptly
at
a
rate
of
1
image/sec,
while
in
Wood

(2013),
the
virtual
objects
moved
smoothly
over
time
through
a
60°
viewpoint
range
at
24

images/sec.
To
test
whether
the
impoverished
visual
stimuli
used
in
the
current

experiment
impaired
chicks’
object
recognition
abilities,
we
compared
performance
in
the

current
study
to
chicks’
performance
in
Wood
(2013).
Figure
19
shows
the
mean

recognition
performance
from
both
studies.
A
independent
samples
t-‐test
showed
that

performance
was
significantly
higher
in
Wood
(2013)
than
in
the
current
study
(t(19.37)
=

2.13,
p
=
0.05).
Thus,
experience
with
smooth,
continuous
object
motion
over
a
larger

viewpoint
range
appears
to
facilitate
the
development
of
invariant
object
recognition.
This

result
is
consistent
with
recent
studies
demonstrating
that
smooth,
continuous
object

75

motion
facilitates
object
recognition
in
newborns
(Wood,
2016;
Wood,
Prasad,
Goldman,
&

Wood,
2016;
Wood
&
Wood,
under
review).

Discussion

In
this
study,
we
examined
whether
newborn
chicks
can
build
invariant
object

representations
from
highly
impoverished
visual
input
(i.e.,
3
images
of
a
single
virtual

object
separated
by
15°
azimuth
rotations).
Impressively,
many
of
the
chicks
successfully

built
an
invariant
object
representation
soon
after
hatching,
which
shows
that
experience

with
a
rich
visual
world
filled
with
diverse
objects
is
not
necessary
for
developing
invariant

Figure 19. Average recognition performance for the present study and for Experiment 1 from
Wood (2013). The same two virtual objects were used in both studies. In the present study, the
virtual objects moved abruptly over time through a 30° viewpoint range at 1 image/second,
whereas in Wood (2013), the virtual objects moved smoothly over time through a 60° viewpoint
range at 24 images/second. Thus, compared with Wood (2013), the chicks in the present study
observed a smaller number of unique images of the object (3 unique images versus 72 unique
images), a smaller range of movement (30° viewpoint range versus 60° viewpoint range), and
unnatural (abrupt) versus natural (smooth) object motion. Performance was significantly above
chance in both studies; however, recognition performance was significantly higher in Wood
(2013) than in the present study. Together, these studies show that it is possible to impair
chicks’ object recognition abilities by presenting highly impoverished visual object input at the
onset of vision.

76

object
recognition.
This
finding
opens
up
largely
unexplored
experimental
avenues
for

probing
the
initial
state
of
invariant
object
recognition
and
charting
how
that
initial
state

changes
as
a
function
of
specific
visual
experiences.

Implications
of
Our
Findings
and
Comparison
with
Previous
Studies

We
have
previously
reported
invariant
object
recognition
in
newborn
chicks
(Wood,

2013;
Wood,
2014a);
the
present
study
extends
this
previous
research
in
five
ways.
First,

these
results
provide
an
existence
proof
that
newborn
chicks
can
build
invariant
object

representations
from
extremely
impoverished
visual
input.
In
previous
studies
(Wood,

2013;
2014a),
chicks
were
shown
objects
that
moved
smoothly
over
time
(24
frames/sec),

thereby
presenting
large
numbers
of
unique
and
gradually
changing
images
of
the
objects.

Conversely,
in
the
present
study,
the
object
animations
were
far
more
sparse
(i.e.,
there

were
only
three
unique
images
of
the
object),
which
interrupted
the
natural
temporal

stability
of
the
visual
object
input
(i.e.,
the
objects
did
not
change
smoothly
over
time).

Thus,
the
chicks
never
observed
their
imprinted
object
(or
any
other
object)
move
with

smooth,
continuous
motion.
Nevertheless,
some
of
the
chicks
were
able
to
build
an

invariant
object
representation
from
this
impoverished
input.
For
these
subjects,
three

unique
images
of
an
object
were
sufficient
input
to
build
an
invariant
object
representation.

Second,
these
results
suggest
that
it
is
possible
to
impair
invariant
object

recognition
in
newborn
chicks
by
presenting
abnormally
patterned
visual
input.
Although

group
performance
was
above
chance,
performance
was
significantly
lower
compared
to

previous
experiments
in
which
the
virtual
object
moved
smoothly
over
time
and
rotated

through
a
larger
viewpoint
range
(Wood,
2013;
see
Figure
19
for
comparison
of

77

performance
between
studies).
Thus,
newborn
visual
systems
appear
to
operate
best
over

a
specific
type
of
patterned
visual
input.
It
would
be
interesting
for
future
studies
to

characterize
the
nature
of
this
‘optimal
space’
of
visual
object
input.

Third,
these
results
indicate
that
invariant
object
recognition
in
newborn
chicks
is

not
subject
to
the
well-‐documented
“viewpoint
effect”
observed
in
studies
of
human
adults

(i.e.,
larger
viewpoint
changes
lead
to
greater
costs
in
object
recognition
performance;

Hayward
&
Williams,
2000;
Tarr,
Williams,
Hayward,
&
Gauthier,
1998).
We
tested
chicks

on
a
wide
range
of
viewpoints,
consisting
of
systematic
15°
changes
in
azimuth
and

elevation
rotations.
This
allowed
us
to
test
whether
larger
viewpoint
changes
are
more

difficult
to
recognize
than
smaller
viewpoint
changes.
We
found
no
significant
differences

in
chicks’
recognition
abilities
across
the
larger
versus
smaller
viewpoint
changes.
Chicks

were
able
to
build
invariant
object
representations
that
generalized
beyond
the
imprinted

viewpoint
range,
but
the
degree
of
generalization
did
not
vary
as
a
function
of
the
degree
of

viewpoint
change.

Importantly,
variation
in
performance
across
different
viewpoints
could
be

accounted
for
by
the
amount
of
self-‐occlusion
produced
by
the
viewpoint
range.

Specifically,
viewpoint
changes
that
resulted
in
higher
levels
of
self-‐occlusion
(regardless
of

the
absolute
change
in
rotation
degree)
were
associated
with
lower
recognition

performance.
When
objects
are
presented
from
viewpoints
with
greater
self-‐occlusion,

critical
2D
or
3D
subfeatures
used
for
recognition
may
not
be
identifiable.

Fourth,
we
demonstrated
that
chicks’
object
recognition
abilities
cannot
be

explained
by
low-‐level
retina-‐like
or
V1-‐like
neuronal
representations.
Prior
experiments

have
confirmed
that
chicks’
object
recognition
abilities
could
not
be
explained
by
overall

78

brightness
(Wood,
2014a)
or
retina-‐like
(pixel-‐wise)
similarity
(Wood,
2013;
Wood,

2014a).
Here,
we
performed
additional
analyses
using
simulated
Gabor
jet
activation
to

measure
the
V1-‐like
similarity
between
the
input
animations
and
the
test
animations.
We

found
that
chicks’
recognition
performance
did
not
vary
as
a
function
of
the
V1-‐like

similarity
between
the
input
and
test
animations.
Further,
we
found
that
neither
a
model

using
pixel-‐like
representations
nor
a
model
using
V1-‐like
representations
was
able
to

successfully
predict
object
identity
in
this
experiment
(Figure
18).
These
results
indicate

that
chicks
build
selective
and
tolerant
object
representations,
akin
to
those
found
in

higher-‐level
cortical
visual
areas
(DiCarlo
et
al.,
2012).

Finally,
our
results
provide
evidence
that
invariant
object
recognition
emerges
in
a

consistent
manner
across
different
newborn
subjects.
The
chicks’
patterns
of
recognition

performance
across
the
individual
viewpoints
were
strongly
correlated
with
one
another

(Figure
16).
This
suggests
that
there
are
constraints
on
the
development
of
invariant
object

recognition
in
newborn
visual
systems.
However,
the
data
also
revealed
substantial

variation
in
chicks’
object
recognition
abilities
(see
Figure
14).
Despite
being
raised
in

identical
visual
environments,
some
chicks
were
able
to
recognize
their
imprinted
object

robustly
across
the
novel
viewpoints,
whereas
other
chicks
were
not.
Future
studies
could

use
this
controlled-‐rearing
method
to
further
examine
both
the
nature
of
the
constraints
on

early
emerging
object
recognition
abilities
and
the
sources
of
the
individual
variation

across
subjects.

In
summary,
the
present
study
provides
evidence
that
the
domestic
chick
is
a

promising
animal
model
for
studying
the
emergence
of
invariant
object
recognition
in
a

newborn
visual
system.
We
have
shown
how
a
fully
automated
controlled-‐rearing

79

technique
can
be
used
to
study
the
initial
state
of
invariant
object
recognition
in
newborn

chicks
with
high
precision.
Thus
far,
our
approach
indicates
that
newborn
neural
circuits

are
surprisingly
powerful,
capable
of
building
invariant
object
representations
from

impoverished
input
at
the
onset
of
vision.

80

Chapter
5:
Face
recognition
in
newborn
chicks
at
the
onset
of
vision

(Corresponding
publication:
Wood,
S.
M.
W.
&
Wood,
J.
N.
(2015)
Face
recognition
in
newly

hatched
chicks
at
the
onset
of
vision.
Journal
of
Experimental
Psychology:
Animal
Learning

and
Cognition,
41(2),
206.)

Abstract

How
does
face
recognition
emerge
in
the
newborn
brain?
To
address
this
question,
we
used

an
automated
controlled-‐rearing
method
with
a
newborn
animal
model:
the
domestic
chick

(Gallus
gallus).
This
automated
method
allowed
us
to
examine
chicks’
face
recognition

abilities
at
the
onset
of
both
face
experience
and
object
experience.
In
the
first
week
of
life,

newborn
chicks
were
raised
in
controlled-‐rearing
chambers
that
contained
no
objects

other
than
a
single
virtual
human
face.
In
the
second
week
of
life,
we
used
an
automated

forced-‐choice
testing
procedure
to
examine
whether
chicks
could
distinguish
that
familiar

face
from
a
variety
of
unfamiliar
faces.
Chicks
successfully
distinguished
the
familiar
face

from
most
of
the
unfamiliar
faces—for
example,
chicks
were
sensitive
to
changes
in
the

face’s
age,
gender,
and
orientation
(upright
versus
inverted).
Thus,
chicks
can
build
an

accurate
representation
of
the
first
face
they
see
in
their
life.
These
results
show
that
the

initial
state
of
face
recognition
is
surprisingly
powerful:
newborn
visual
systems
can
begin

encoding
and
recognizing
faces
at
the
onset
of
vision.

81

Introduction

Social
animals
depend
heavily
on
their
ability
to
recognize
faces.
For
instance,
face

recognition
(i.e.,
the
ability
to
encode
and
recognize
specific
faces)
allows
animals
to
form

and
maintain
social
relationships
and
identify
key
competitors
in
their
group.
Previous

studies
have
examined
face
recognition
abilities
days,
months,
and
years
after
birth
(e.g.,

Carey
&
Diamond,
1977;
de
Haan,
Johnson,
Maurer,
&
Perrett,
2001;
Frank,
Vul,
&
Johnson,

2009;
Kelly
et
al.,
2007;
Pascalis,
Deschonen,
Morton,
Deruelle,
&
Fabregrenet,
1995;
Sugita,

2008).
To
date,
however,
little
is
known
about
the
‘initial
state’
of
face
recognition
(i.e.,
the

state
of
face
recognition
at
the
onset
of
vision).
Can
newborn
9

animals
encode
and

recognize
faces
at
the
onset
of
both
face
experience
and
object
experience?
Or
does
face

recognition
have
a
protracted
development,
requiring
extensive
exposure
to
faces
and/or

objects
in
order
to
develop?

In
the
present
study,
we
used
an
automated
controlled-‐rearing
method
to
address

three
questions:
(1)
Can
newborn
animals
build
an
accurate
representation
of
the
first
face

they
see
in
their
life?
(2)
What
types
of
face
changes
can
newborn
animals
detect
at
the

onset
of
vision?
(3)
Are
there
individual
differences
in
newborn
animals’
face
recognition

abilities?

To
be
clear
at
the
outset,
this
study
was
not
designed
to
test
whether
face

recognition
depends
on
the
same
mechanisms
or
different
mechanisms
than
object

recognition.
Rather,
our
goal
was
to
examine
whether
newborn
animals
are
capable
of

encoding
and
recognizing
faces
at
the
onset
of
vision.
In
the
discussion,
we
return
to
the

9
The term “newborn” is used to refer to an animal at the beginning of the post-embryonic phase
of their life cycle.

82

issue
of
whether
face
recognition
and
object
recognition
depend
on
shared
versus
unique

mechanisms
at
the
onset
of
vision.

Chickens
as
an
animal
model
for
studying
the
initial
state
of
face
recognition

Face
recognition
is
a
form
of
visual
learning.
According
to
a
growing
body
of
work
in

the
neurosciences,
visual
learning
occurs
rapidly
within
the
visual
system
(e.g.,
DiCarlo
et

al.,
2012;
Espinosa
&
Stryker,
2012;
Gavornik
&
Bear,
2014).
For
instance,
the
visual
cortex

uses
statistical
redundancies
present
in
the
natural
environment
to
fine-‐tune
the
response

properties
of
neurons
(Edelman
&
Intrator,
2003;
Olshausen
&
Field,
1996).
Further,

studies
of
monkeys
show
that
category-‐selective
regions
emerge
in
the
cortex
on
the
basis

of
early
visual
experience
(Srihasam,
Mandeville,
Morocz,
Sullivan,
&
Livingstone,
2012),

with
significant
changes
in
the
response
patterns
of
neural
populations
occurring
as
little

as
1
hour
after
exposure
to
an
altered
visual
world
(Li
&
DiCarlo,
2008).
Since
the
visual

system
is
rapidly
shaped
by
visual
experience,
studying
the
origins
of
a
visual
learning

ability
like
face
recognition
requires
a
controlled-‐rearing
approach
with
a
newborn
animal

model.
With
controlled-‐rearing
methods,
it
is
possible
to
systematically
manipulate
an

animal’s
visual
experiences,
and
thus,
assess
the
impact
of
specific
experiences
on
the

development
of
perceptual
and
cognitive
abilities.

In
the
current
study,
we
used
a
controlled-‐rearing
method
with
domestic
chicks

(Gallus
gallus).
Five
characteristics
make
chicks
an
ideal
animal
model
for
studying
the

initial
state
of
face
recognition.
First,
chicks
can
be
raised
in
environments
devoid
of
both

faces
and
objects.
Unlike
newborn
primates,
newborn
chicks
do
not
require
parental
care

and,
because
of
precocial
motor
development,
are
immediately
able
to
explore
their

83

environment.
Second,
chicks
imprint
to
conspicuous
objects
experienced
in
the
first
few

days
of
life
(e.g.,
Bateson,
2000;
Horn,
2004).
Chicks
develop
a
strong
attachment
to
their

imprinted
objects,
treating
them
as
social
partners.
Thus,
this
imprinting
behavior
can
be

used
to
test
chicks’
visual
recognition
abilities
without
training.
Third,
adult
birds
can

discriminate
between
human
faces,
and
rely
on
similar
facial
features
for
face
recognition

as
human
adults
(Bogale,
Aoyama,
&
Sugita,
2011;
Gibson,
Wasserman,
Gosselin,
&
Schyns,

2005).
These
findings
provide
evidence
that
human
and
avian
visual
systems
build
similar

face
representations
as
one
another.
Fourth,
chicks
show
a
preference
for
face-‐like
stimuli

at
the
onset
of
face
experience
(Rosa-‐Salva,
Farroni,
Regolin,
Vallortigara,
&
Johnson,
2011;

Rosa-‐Salva
et
al.,
2010;
Rosa-‐Salva,
Regolin,
&
Vallortigara,
2012),
akin
to
newborn

humans.
The
current
study
builds
on
these
findings
by
examining
whether
chicks
can

encode
and
recognize
specific
faces
at
the
onset
of
vision.
Fifth,
chickens
and
humans
use

similar
neural
circuits
to
process
sensory
information
(Karten,
2013).
Although

mammalian
and
avian
brains
differ
in
their
macroarchitecture
(i.e.,
layered
versus
nuclear

organization,
respectively),
they
are
nearly
identical
from
the
perspective
of
the
cells
and

circuits
that
process
sensory
information
(reviewed
by
Karten,
2013).
Together,
these

characteristics
make
chicks
an
ideal
and
unique
animal
model
for
studying
the
emergence

of
face
recognition
in
a
biological
visual
system.

Previous
controlled-‐rearing
studies
have
also
demonstrated
that
chicks
are
a

promising
animal
model
for
studying
the
origins
of
object
recognition
and
visual
learning

more
generally.
For
instance,
chicks
begin
binding
color
and
shape
features
into
integrated

object
representations
at
the
onset
of
vision
(Wood,
2014)
and
can
build
a
view-‐invariant

representation
of
the
first
object
they
see
in
their
life
(Wood,
2013,
2015).
Chicks
also

84

begin
encoding
and
recognizing
movements
and
movement
sequences
within
the
first
few

days
of
life
(Goldman
&
Wood,
2015).
The
present
study
builds
on
this
previous
work
by

examining
whether
newborn
chicks
can
build
accurate
representations
of
faces
at
the
onset

of
vision.
Face
recognition
is
a
prototypical
example
of
subordinate-‐level
object
recognition

because
all
faces
share
a
general
configuration
(Carey,
1992).
Thus,
face
recognition

requires
more
fine-‐grained
discrimination
than
basic-‐level
object
recognition.

An
automated
controlled-‐rearing
method
for
studying
face
recognition

In
the
past,
newborn
animals’
behavior
has
been
quantified
through
direct

observation
by
trained
researchers.
While
direct
observation
has
revealed
many
important

insights
about
newborn
cognition,
there
are
limitations
to
this
approach
(Dell
et
al.,
2014).

Direct
observation
produces
a
limited
amount
of
data
with
relatively
low
spatial
and

temporal
resolution.
Further,
the
resulting
data
is
a
subjective
description
of
the
subject’s

behavior,
rather
than
an
exact
record
of
events.
Direct
observation
therefore
allows
for
the

possibility
of
experimenter
bias,
a
well-‐recognized
problem
in
both
comparative
and

developmental
psychology.
In
contrast
to
direct
observation,
automated
experimental

methods
allow
researchers
to
collect
large
amounts
of
data
from
each
subject
and
quantify

behavior
at
scales
not
previously
possible.
Further,
since
the
observations
are
not
made
by

a
researcher,
automated
methods
remove
the
possibility
of
experimenter
bias.

Here,
we
describe
an
automated
‘complete
data’
controlled-‐rearing
method
for

studying
the
initial
state
of
face
recognition.
This
automated
approach
has
previously
been

used
to
study
the
initial
state
of
object
recognition
(Wood,
2013,
2014,
2015)
and
action

recognition
(Goldman
&
Wood,
2015);
here,
we
extend
the
method
to
the
domain
of
face

85

recognition.
We
use
the
term
‘complete
data’
because
the
method
involves
recording
all
of

the
newborn
subjects’
behavior
(24
hours/day,
7
days/week)
with
high
precision
(9

samples/second).
This
approach
produces
a
complete
digital
record
of
each
subject’s

behavior
across
their
lifespan.

The
goal
of
the
current
study
was
to
examine
the
initial
state
of
face
recognition
by

testing
newborn
chicks’
face
recognition
abilities
across
a
wide
range
of
face-‐change

conditions.
In
their
first
week
of
life
(the
input
phase),
chicks
were
raised
in
controlled-‐
rearing
chambers
that
contained
no
objects
other
than
a
single
virtual
human
face.
In
their

second
week
of
life
(the
test
phase),
we
tested
whether
chicks
could
distinguish
that
virtual

face
from
a
variety
of
unfamiliar
faces.
Since
we
recorded
all
of
the
chicks’
behavior,
it
was

possible
to
present
each
subject
with
a
large
number
of
test
trials
(~140
test
trials
per

chick)
across
10
face-‐change
conditions.
As
a
result,
we
were
able
to
determine
the
features

used
by
each
chick
to
recognize
faces
and
compare
the
face
recognition
strategies

employed
by
different
subjects.

Methods

Thirteen
domestic
chicks
of
unknown
sex
were
tested.
No
subjects
were
excluded

from
the
analyses.
The
subjects
were
tested
in
the
controlled-‐rearing
chambers
described

in
Chapters
2-‐4.

During
the
input
phase
(the
first
week
of
life),
chicks
were
raised
in
an
environment

that
contained
a
single
virtual
face.
Imprinting
in
chickens
is
subject
to
a
critical
period,

which
ends
approximately
three
days
after
hatching.
Thus,
to
ensure
that
the
chicks
had

fully
imprinted
to
the
virtual
face,
we
exposed
the
chicks
to
the
virtual
face
for
the
first

86

seven
days
of
life.
Six
chicks
were
shown
an
older
male
face
(ear-‐to-‐ear
width
=
6.2
cm,

height
=
10
cm)
and
seven
chicks
were
shown
a
younger
female
face
(ear-‐to-‐ear
width
=
6.5

cm;
height
=
10
cm)
(Figure
20).
The
virtual
face
moved
continuously,
rotating
smoothly

Figure 20. (A) An illustration of the controlled-rearing chambers from a bird’s eye perspective
(not shown to scale). (B) A schematic of the presentation schedule of the virtual faces on the two
display walls during the input phase (top) and the test phase (bottom). (C) Images from the male
face animation shown during the input phase. (D) Images from the female face animation shown
during the input phase. Each chick was shown either the male face or the female face.

87

through
a
180°
viewpoint
range
about
a
frontoparallel
vertical
axis
passing
through
its

centroid.
The
animations
contained
24
frames/second.
The
individual
face
frames
were

created
using
FaceGen
software
(Singular
Inversions
Inc.).
The
faces
were
displayed
on
a

uniform
white
background
and
positioned
1
cm
off
the
ground
in
the
middle
of
the
display

walls.
The
imprinted
face
appeared
for
an
equal
amount
of
time
on
the
left
and
right
display

wall
and
switched
walls
every
two
hours,
following
a
one-‐minute
period
of
darkness

(Figure
20B).

We
used
human
faces
(rather
than
chicken
faces)
because
the
face
images
could
be

precisely
manipulated
along
a
variety
of
dimensions
using
the
FaceGen
software.
More

importantly,
using
human
faces
allows
for
a
more
direct
comparison
with
studies
of
face

recognition
in
humans
and
other
avian
species
(Bogale
et
al.,
2011;
Gibson
et
al.,
2005;

Troje,
Huber,
Loidolt,
Aust,
&
Fieder,
1999).

Test
Phase

During
the
test
phase
(the
second
week
of
life),
we
probed
the
informational
content

of
the
face
representation
built
by
each
chick
by
using
an
automated
two-‐alternative
forced

choice
testing
procedure.
On
each
test
trial,
the
imprinted
face
was
projected
onto
one

display
wall
and
an
unfamiliar
face
was
projected
onto
the
other
display
wall
(see
Figure

20A).
If
chicks
recognized
their
imprinted
face,
then
they
should
have
spent
a
greater

proportion
of
time
in
proximity
to
the
imprinted
face
compared
to
the
unfamiliar
face

during
these
test
trials.
The
unfamiliar
faces
had
the
same
size,
motion
speed,
and

viewpoint
range
as
the
imprinted
face.
The
test
trials
lasted
24
minutes
and
were
separated

from
one
another
by
46-‐minute
rest
periods.
During
the
rest
periods,
the
imprinted
face

88

appeared
on
one
display
wall
and
a
white
screen
appeared
on
the
other
display
wall.
Each

chick
received
20
test
trials
per
day
(two
test
trials
for
each
of
the
10
face-‐change

conditions
described
below).
The
conditions
were
presented
in
randomized
blocks

throughout
the
test
phase.

Since
this
was
the
first
study
to
examine
chicks’
face
recognition
abilities
at
the

onset
of
vision,
we
presented
subjects
with
a
wide
range
of
face
change
conditions
to
obtain

a
general
sense
of
their
recognition
abilities
(Figure
21).
In
the
“Edges
Only”
condition,
the

unfamiliar
face
was
a
line
drawing
of
the
imprinted
face.
In
the
“No
Color”
condition,
the

unfamiliar
face
was
created
by
removing
all
color
information
from
the
imprinted
face.
We

included
these
two
conditions
to
test
whether
chicks
encode
only
the
edge/shape
features

of
a
face
or
whether
they
also
encode
the
color
features
of
a
face.
In
the
“Features
Only”

condition,
the
unfamiliar
face
had
the
eyes
and
mouth
of
the
imprinted
face
but
without

any
of
the
surrounding
facial
context.
In
the
“Repositioned
Features”
condition,
the

unfamiliar
face
was
created
by
moving
the
facial
features
of
the
imprinted
face
to
new

positions.
We
included
these
two
conditions
to
test
whether
chicks
encode
only
the
eye
and

mouth
features
of
a
face
or
whether
they
encode
the
surrounding
facial
context,
and
also
to

examine
whether
chicks
encode
the
positions
of
the
eyes
and
mouth
within
the
face.
In
the

“Inverted”
condition,
the
unfamiliar
face
was
identical
to
the
imprinted
face,
but
in
an

inverted
position.
We
tested
chicks
in
this
condition
to
examine
whether
they
encode
the

spatial
orientation
of
a
face.
In
the
“Different
Age”
condition,
we
changed
the
age
of
the

imprinted
face
(i.e.,
for
the
young
woman
imprinted
face,
the
unfamiliar
face
was
an
older

woman;
and
for
the
older
man
imprinted
face,
the
unfamiliar
face
was
a
younger
man).
We

modified
the
gender
of
the
imprinted
face
in
two
conditions.
In
the
“Different
Gender

89

Coloring”
condition,
the
unfamiliar
face
had
the
same
shape
as
the
imprinted
face,
but
with

color
features
that
were
more
characteristic
of
the
opposite
gender.
In
the
“Different

Gender
Shape”
condition,
the
unfamiliar
face
had
the
same
color
as
the
imprinted
face,
but

with
shape
features
that
were
more
characteristic
of
the
opposite
gender.
We
tested
chicks

in
these
conditions
to
examine
whether
they
can
distinguish
between
faces
of
different

identities
based
on
gender
and
age
information.
Finally,
we
tested
chicks’
sensitivity
to

facial
expressions:
the
unfamiliar
face
was
identical
to
the
imprinted
face
except
that
it
had

either
an
angry
(“Angry
Expression”
condition)
or
fearful
(“Fearful
Expression”
condition)

expression.
We
tested
chicks
in
these
conditions
to
examine
whether
they
build

representations
of
faces
that
are
tolerant
to
changes
in
expression.

Results

To
compute
each
chick’s
recognition
performance,
we
first
computed
the
proportion

of
time
each
chick
spent
with
the
correct
animation
compared
to
the
incorrect
animation

for
the
test
trials
in
which
the
imprinted
face
switched
display
walls
after
the
rest
period

and
for
the
test
trials
in
which
the
imprinted
face
stayed
on
the
same
display
wall
after
the

rest
period.
We
then
computed
the
average
of
these
two
values
to
obtain
a
single

recognition
performance
score
for
each
chick
in
the
condition.
10

10
Please see the original published data (Wood & Wood, 2015) for a Bayesian analysis of the
data. Here, I have re-analyzed the data using traditional null hypothesis testing to match the other
studies presented in Chapters 2, 3, 4, and 6.

90

Results
are
shown
in
Figure
21.
A
repeated
measures
ANOVA
with
the
within-‐
subjects
factor
of
condition
and
the
between-‐subjects
factor
of
imprinted
face
revealed
a

significant
main
effect
of
condition
(F(9,
99)
=
14.505,
p
<
10
-‐13
),
but
no
significant
main

effect
of
imprinted
face
or
the
interaction
between
condition
and
imprinted
face.
Chicks

were
able
to
recognize
their
familiar
face
in
7
of
the
10
conditions:
Edges
Only
(t(12)
=

22.418,
p
<
10
-‐10
),
No
Color
(t(12)
=
22.418,
p
<
10
-‐9
),
Features
Only
(t(12)
=
6.274,
p
<
10
-‐
4
),
Different
Gender
Coloring
(t(12)
=
4.648,
p
<
.001),

Different
Gender
Shape
(t(12)
=

Figure 21. Results from the 10 face-change conditions. Each bar shows the average percent
of correct trials in each condition for the male (blue bars) and female (green bars) imprinted
face. Chance performance was 50%. Error bars denote standard error.

91

2.283,
p
=
.041)
11
,
Inverted
(t(12)
=
8.348,
p
<
10
-‐5
),
and
Different
Age
conditions
(t(12)
=

6.593,
p
<
10
-‐4
).
Performance
was
not
above
chance
levels
in
the
Fearful
Expression

condition,
Angry
Expression
condition,
or
Repositioned
Features
condition
(all
Ps
>
.05).

Analysis
of
Effect
Sizes

To
quantify
the
magnitude
of
the
chicks’
performance,
we
computed
a
one
sample

Cohen’s
d
for
each
condition.
We
found
large
effect
sizes
(i.e.,
greater
than
0.8)
for
6
of
the

10
conditions:
Edges
Only
(d
=
6.2),
No
Color
(d
=
5.2),
Features
Only
(d
=
1.7),
Different

Gender
Coloring
(d
=
1.3),
Inverted
(d
=
2.3),
and
Different
Age
(d
=
1.8).
We
also
found
a

medium
effect
size
for
the
Different
Gender
Shape
condition
(d
=
0.6).

Analysis
of
Change
in
Performance
Over
Time

To
compute
whether
there
was
any
difference
in
performance
depending
on
trial

day,
we
performed
a
repeated
measures
ANOVA
with
the
within-‐subjects
factor
of
trial
day.

The
ANOVA
revealed
a
significant
main
effect
of
trial
day
(F(6,
72)
=
6.034,
p
<
10
-‐4
).
As

shown
in
Figure
22,
performance
on
Day
1
was
significantly
lower
than
all
other
trial
days

(paired
t-‐tests,
all
Ps
<
.01).
Notably,
on
Days
2-‐7
of
testing,
performance
was
significantly

above
chance
on
8
of
the
10
conditions
(Edges
Only,
No
Color,
Features
Only,
Different

Gender
Coloring,
Different
Gender
Shape,
Inverted,
Different
Age,
and
Angry
Face

conditions;
all
Ps
<
.05;
all
conditions
surviving
Holm-‐Bonferroni
correction).
Thus,

additional
research
is
necessary
to
determine
the
extent
to
which
chicks’
face
recognition

abilities
improve
over
time
in
these
impoverished
visual
environments.

11
However, note that after applying a Holm-Bonferroni correction, performance in the Different
Gender Shape condition was no longer statistically significant.

92

Analysis
of
Individual
Subject
Performance

With
this
controlled-‐rearing
method,
we
were
able
to
collect
a
large
number
of
test

trials
from
each
chick.
This
made
it
possible
to
measure
each
chick’s
face
recognition

abilities
with
high
precision.
First,
we
examined
whether
all
of
the
chicks
were
able
to
build

an
accurate
representation
of
their
imprinted
face,
by
computing
whether
each
chick’s

performance
across
the
test
trials
exceeded
chance
level.
As
shown
in
Figure
23,
12
of
the

13
chicks
spent
more
time
with
the
imprinted
face
compared
to
the
unfamiliar
face
on
the

test
trials
(one-‐sample
t-‐tests,
all
Ps
<
.001).
This
result
indicates
that
almost
all
of
the

chicks
were
able
to
build
an
accurate
representation
of
the
virtual
face.

Figure 22. Change over time results. The graph illustrates group mean performance
over the full set of face change conditions shown during the test phase, computed for
the first, second, third, etc., test day. Chance performance was 50%. Error bars denote
standard error.

40%$
45%$
50%$
55%$
60%$
65%$
70%$
75%$
80%$
1$ 2$ 3$ 4$ 5$ 6$ 7$
Propor%on'of'%me'with'Imprinted'Face'
versus'Unfamiliar'Face'
Day'

93

Second,
we
examined
whether
the
chicks
used
the
same
general
strategy
as
one

another
to
distinguish
the
imprinted
face
from
the
unfamiliar
faces.
Figure
23B
shows
each

chick’s
sensitivity
to
each
of
the
face
changes.
Visual
inspection
of
Figure
23B
shows
that

the
majority
of
the
chicks
were
sensitive
to
the
same
face
changes.
To
examine
whether
the

Figure
23.
(A)
Performance
of
each
individual
subject
(ordered
by
performance).
The

graph
shows
the
total
number
of
correct
and
incorrect
test
trials
for
each
chick
across

the
test
phase.
P-‐values
denote
the
statistical
difference
between
the
number
of
correct

and
incorrect
test
trials
(computed
through
one-‐tailed
binomial
tests).
(B)
The

percentage
of
correct
trials
for
each
chick
in
each
condition.
Chance
performance
was

50%.
Subjects
are
ordered
by
overall
performance
for
each
imprinted
face.
(C)
A

correlation
matrix
showing
the
correlation
in
face
recognition
performance
for
each
pair

of
chicks.
Each
box
shows
the
correlation
between
two
chicks’
percent
of
successful

trials
in
each
condition.
The
subjects
are
ordered
by
overall
performance
for
each

imprinted
face.
The
cells
are
color-‐coded
by
correlation
value:
green
values
=
positive

correlation
in
performance;
red
values
=
negative
correlation
in
performance.
The
color

scale
reflects
the
full
range
of
possible
correlation
values.

94

chicks’
face
recognition
abilities
were
correlated
with
one
another,
we
created
a
correlation

matrix
(Figure
23C).
This
matrix
shows
the
correlation
in
face
recognition
performance
for

each
pair
of
chicks
across
the
conditions
(i.e.,
each
box
shows
the
correlation
between
two

chicks’
percent
of
correct
trials
in
each
condition).
Green
values
indicate
a
positive

correlation
in
performance,
red
values
indicate
a
negative
correlation
in
performance,
and

yellow
values
indicate
a
weak
correlation.
Chicks’
face
recognition
abilities
were
highly

correlated
across
the
conditions,
with
an
average
between-‐subject
correlation
of
r
=
.58

(SEM
=
0.02).

Analysis
of
Stimuli
Features

What
visual
features
are
the
chicks
using
to
recognize
their
imprinted
face?

Research
on
human
subjects
has
found
that
adults
use
Gabor
Jet
features
to
recognize
faces

(Yue,
Biederman,
Mangini,
von
der
Malsburg,
&
Amir,
2012).
To
determine
whether
the

chicks
also
rely
on
Gabor
Jet
features,
we
computed
Gabor
Jet
dissimilarity
(using
the

online
applet
provided
at
http://geon.usc.edu/GJW/)
between
the
imprinted
face
and
the

test
faces.
For
each
animation,
the
applet
computed
the
Gabor
Jet
magnitudes
for
the
frame

of
the
animation
in
which
the
face
is
at
a
0°
angle
to
the
viewer.
Finally,
the
applet

computed
the
Euclidean
distance
between
the
imprinted
faces
and
each
of
their
respective

test
faces.
Overall,
performance
was
not
correlated
with
Gabor
Jet
dissimilarity

(Spearman’s
r
=
-‐.029,
p
=
.905).
Thus,
newborn
chicks
do
not
appear
to
rely
on
the
same

features
as
adult
humans
to
recognize
faces.

95

Discussion

This
study
examined
whether
newborn
chicks
can
encode
and
recognize
faces
at
the

onset
of
vision.
Specifically,
chicks
were
raised
in
automated
controlled-‐rearing
chambers

that
recorded
all
of
their
behavior
with
high
precision.
In
their
first
week
of
life,
chicks’

visual
experience
with
faces
and
objects
was
limited
to
a
single
virtual
face
rotating
around

a
single
axis.
In
their
second
week
of
life,
we
tested
whether
chicks
could
distinguish
that

virtual
face
from
a
variety
of
unfamiliar
faces.
Three
main
findings
emerged.

First,
despite
lacking
any
prior
face
and
object
experience,
chicks
were
able
to
build

an
accurate
representation
that
supported
face
recognition
across
a
range
of
conditions.

While
previous
studies
have
shown
that
newborn
animals
can
detect
face-‐like

configurations
soon
after
birth
(Johnson,
Dziurawiec,
Ellis,
&
Morton,
1991;
Rosa-‐Salva
et

al.,
2011;
Rosa-‐Salva
et
al.,
2010),
the
current
study
indicates
that
newborn
animals
can

also
encode
and
recognize
specific
faces
at
the
onset
of
vision.
For
instance,
chicks
were

sensitive
to
changes
in
their
imprinted
face’s
age,
gender,
and
orientation
(upright
vs.

inverted).
Further,
chicks
showed
little
to
no
sensitivity
to
changes
in
facial
expression,

which
suggests
that
a
chick’s
first
face
representation
can
be
tolerant
to
some
identity-‐
preserving
facial
transformations.
Together,
this
pattern
of
results
shows
that
chicks
can

build
a
selective
and
tolerant
representation
of
a
face.
This
study
extends
the
existing

literature
concerning
chicks’
visual
learning
abilities.
Previous
controlled-‐rearing

experiments
show
that
chicks
can
build
an
integrated
and
invariant
representation
of
the

first
object
they
see
in
their
environment
(Wood,
2013,
2014).
The
present
study
shows

that
chicks
can
also
build
an
accurate
representation
of
the
first
face
they
see.
Thus,
chicks

can
learn
rapidly
about
a
variety
of
entities
at
the
onset
of
vision.

96

Second,
these
results
provide
evidence
that
chicks
build
similar
face
representations

as
one
another
at
the
onset
of
face
and
object
experience.
As
shown
in
Figure
23B,
most
of

the
chicks
were
sensitive
to
the
same
visual
features
when
recognizing
faces,
and
as
shown

in
Figure
23C,
most
of
the
chicks’
face
recognition
abilities
were
highly
correlated
with
one

another.
Thus,
different
chicks
use
a
common
strategy
to
distinguish
between
faces.

Third,
these
results
begin
to
reveal
the
types
of
face
information
that
can
be
encoded

at
the
onset
of
vision.
Our
results
provide
evidence
that
color
information
is
an
important

cue
for
chicks’
face
recognition
abilities
because
subjects
reliably
distinguished
their

imprinted
face
from
unfamiliar
faces
that
had
different
color
features,
but
identical
shape

features
(i.e.,
Edges
Only,
No
Color,
and
Different
Gender
Coloring
conditions).
Likewise,

many
studies
have
shown
that
color
information
plays
an
important
role
in
human
adults’

face
recognition
abilities
(e.g.,
Farah,
Wilson,
Drain,
&
Tanaka,
1998;
Hill,
Bruce,
&

Akamatsu,
1995;
Said
&
Todorov,
2011).
Our
results
also
provide
suggestive
evidence
that

chicks
use
shape/position
information
to
recognize
faces,
because
subjects
reliably

distinguished
the
imprinted
face
from
an
inverted
version
of
the
imprinted
face.
More

generally,
these
results
accord
with
previous
controlled
rearing
experiments
of
object

recognition,
which
show
that
chicks
can
encode
both
the
color
and
shape
of
objects
(Wood,

2014).

While
the
current
study
focused
on
the
initial
state
of
face
recognition,
previous

developmental
studies
have
shown
that
experience
and
maturation
play
an
important
role

in
shaping
and
calibrating
face
recognition
machinery,
with
significant
changes
occurring

over
the
first
16
years
of
life
in
humans
(Bruce
et
al.,
2000;
Carey
&
Diamond,
1977;

Mondloch,
Le
Grand,
&
Maurer,
2010).
Some
researchers
have
suggested
that
the

97

development
of
face
recognition
is
protracted
because
sensitivity
to
configural
effects
does

not
emerge
until
relatively
late
in
development
(Carey
&
Diamond,
1977).
Our
findings
are

consistent
with
this
suggestion
because
chicks
were
not
able
to
distinguish
their
imprinted

face
from
an
unfamiliar
face
in
which
the
features
of
the
imprinted
face
were
located
at

different
positions
(i.e.,
Repositioned
Features
condition).

It
is
important
to
emphasize
two
potential
limitations
of
the
current
study.
First,

these
chicks
observed
the
imprinted
face
for
an
extended
period
of
time
throughout
the

input
phase.
Thus,
additional
studies
are
needed
to
determine
whether
chicks
can
build
an

accurate
face
representation
after
seeing
a
face
briefly,
akin
to
human
adults,
or
whether

they
need
to
see
a
face
for
an
extended
period
of
time.

Second,
this
experiment
was
not
designed
to
test
whether
chicks’
face
recognition

abilities
depend
on
domain-‐specific
versus
domain-‐general
recognition
mechanisms.
Some

researchers
have
proposed
that
face
recognition
and
object
recognition
depend
on

separate,
domain-‐specific
systems
from
birth
(Carey,
2009;
Spelke
&
Kinzler,
2007;

Vallortigara,
2012).
Conversely,
other
researchers
have
proposed
that
face
recognition
and

object
recognition
initially
depend
on
common
domain-‐general
computations,
with

domain-‐specific
neural
populations
emerging
in
the
cortex
on
the
basis
of
visual

experience.
According
to
this
second
proposal,
domain-‐specific
face
recognition
should

emerge
relatively
late
in
development,
only
after
the
animal
has
been
exposed
to
different

classes
of
objects
and
faces
(reviewed
by
Wallis,
2013).
Support
for
this
domain-‐general

position
comes
from
studies
showing
that
face
memory
undergoes
domain-‐specific

development
during
the
first
10
years
of
human
life
(Weigelt
et
al.,
2014),
that
newborns’

early-‐emerging
face
preferences
can
be
explained
by
domain-‐general
computations

98

operating
over
binocular
input
(Wilkinson,
Paikan,
Gredeback,
Rea,
&
Metta,
2014),
and

that
category-‐selective
regions
(e.g.,
regions
selective
for
faces
or
letter
symbols)
emerge
in

the
cortex
on
the
basis
of
early
visual
experiences
(Roder,
Ley,
Shenoy,
Kekunnaya,
&

Bottari,
2013;
Srihasam
et
al.,
2012).
It
would
be
interesting
for
future
studies
to
use
this

automated
controlled-‐rearing
method
to
examine
whether
face
recognition
and
object

recognition
depend
on
shared
versus
unique
computations
at
the
onset
of
vision,
by

examining
whether
newborn
chicks
use
similar
computations
when
building
their
first
face

and
object
representations.

Future
studies
could
also
use
this
controlled-‐rearing
approach
to
explore
a
range
of

other
questions
about
the
initial
state
of
face
recognition.
For
example,
what
facial
features

do
newborn
animals
use
to
recognize
faces
at
the
onset
of
vision?
How
do
these
features

change
as
the
animal
acquires
experiences
with
faces
and/or
objects?
Are
some
face

changes
easier
to
detect
on
male
faces
versus
female
faces?
And
how
do
more
abstract

facial
categories
(e.g.,
categories
for
race,
gender,
and
age)
emerge
in
the
visual
system
as
a

function
of
specific
face
and
object
experiences?

In
sum,
our
study
provides
systematic
evidence
that
newborn
chicks
are
capable
of

recognizing
faces.
Impressively,
chicks
are
able
to
distinguish
different
faces
from
one

another
soon
after
hatching,
which
shows
that
experience
with
a
rich
visual
world
is
not

necessary
for
developing
face
recognition.

99

Chapter
6:
A
slowness
constraint
on
the
development
of
view-‐invariant
face
recognition

Abstract

The
ability
to
recognize
faces
is
central
to
social
behavior.
To
date,
however,
little
is
known

about
how
invariant
face
recognition
emerges
in
the
newborn
brain.
Can
newborns
begin

building
invariant
face
representations
at
the
onset
of
vision?
If
so,
does
the
development

of
this
ability
require
a
particular
type
of
visual
experience
with
faces?
To
address
these

questions,
we
used
an
automated
controlled-‐rearing
method
with
newborn
chicks.
In
the

first
week
of
life,
we
raised
chicks
with
a
single
virtual
face
rotating
through
a
single

viewpoint
range.
In
the
second
week
of
life,
we
tested
whether
the
chicks
built
view-‐
invariant
face
representations.
We
found
that
newborn
chicks
successfully
recognized
the

familiar
face
across
novel
viewpoints.
Moreover,
we
found
that
this
ability
was
impaired

when
newborn
chicks
were
raised
with
a
face
that
moved
quickly
over
time.
Thus,
the

development
of
view-‐invariant
face
recognition
is
subject
to
a
‘slowness
constraint.’
These

results
indicate
that
invariant
face
recognition
can
emerge
rapidly
in
newborns
and
that

the
development
of
this
ability
requires
visual
experience
with
slowly
moving
faces,
akin
to

the
development
of
object
recognition.

100

Introduction

The
ability
to
recognize
faces
quickly
and
accurately
is
critical
to
social
interactions.

Unlike
inanimate
objects,
faces
(and
other
body
parts)
are
self-‐propelled,
and
thus
uniquely

dynamic.
Thus,
like
objects,
faces
in
natural
settings
appear
across
tremendous
variation
in

viewing
situations
(e.g.,
changes
in
facial
expression,
viewpoint,
background,
and
occluding

objects).
The
ability
to
recognize
faces
across
these
identity-‐preserving
image

transformations
is
a
computationally
complex
task
known
as
“invariant
face
recognition”

(Hasselmo,
Rolls,
Baylis,
&
Nalwa,
1989;
Hill,
Schyns,
&
Akamatsu,
1997;
Moses,
Ullman,
&

Edelman,
1996;
Wallis
&
Rolls,
1997).
Since
each
encounter
with
a
face
is
almost
entirely

unique
(in
terms
of
the
image
projected
on
the
retina),
the
visual
system
must
link
these

different
retinal
patterns
to
the
same
face
stored
in
memory.
While
numerous
studies
have

investigated
face
recognition
in
adults,
little
is
known
about
the
origins
of
invariant
face

recognition.

Due
to
challenges
associated
with
testing
newborns
experimentally,
it
has
generally

not
been
possible
to
study
the
initial
state
of
face
recognition
(i.e.,
the
state
of
face

recognition
machinery
at
the
onset
of
vision).
Two
major
limitations
have
hindered

progress.
First,
most
newborn
animals
cannot
be
raised
in
controlled
environments
from

birth.
This
limitation
has
prevented
researchers
from
studying
how
specific
visual

experiences
shape
the
initial
state
of
face
recognition.
Second,
researchers
can
typically

collect
only
a
few
test
trials
from
each
newborn
subject.
This
limitation
has
prevented

researchers
from
obtaining
precise
measurements
of
face
recognition
in
newborns.

Using
automated
controlled
rearing
to
explore
the
origins
of
face
recognition

101

In
previous
studies
of
object
recognition,
Wood
(2013)
addressed
these
limitations

by
using
an
automated
controlled-‐rearing
technique.
In
particular,
Wood
(2013)
was
able

to
study
the
initial
state
of
object
recognition
by
raising
newborn
chicks
in
automated

controlled-‐rearing
chambers
that
provided
complete
control
over
all
visual
object

experiences.
The
automated
chambers
used
computers
to
perform
all
stimuli
presentation

and
data
collection.
Unlike
studies
that
have
observed
newborn
birds
for
5-‐10
minutes

(e.g.,
Martinho
&
Kacelnik,
2016;
Mascalzoni
et
al.,
2010;
Regolin
et
al.,
2011;
Regolin
&

Vallortigara,
1995;
Rosa-‐Salva
et
al.,
2016;
Rosa-‐Salva
et
al.,
2010;
Vallortigara
et
al.,
2005),

this
method
enables
researchers
to
measure
a
newborn’s
first
visual
representation
with

high
precision,
by
collecting
thousands
of
minutes
of
test
data
during
the
experiment.

This
methodology
has
been
fruitful
for
examining
the
developmental
origins
of

object
recognition.
For
example,
studies
using
this
automated
method
have
found
that

newborn
chicks
are
able
to
build
view-‐invariant
representations
of
objects
(Wood,
2013;

Wood
&
Wood,
2015a).
However,
the
development
of
this
ability
is
subject
to
a
“slowness

constraint.”
In
order
for
newborn
chicks
to
develop
view-‐invariant
object
recognition,
they

must
be
raised
in
environments
containing
slowly
moving
objects
(Wood
&
Wood,
2016a).

Specifically,
the
information
content
of
chicks’
object
representations
(i.e.,
viewpoint-‐
specific
and
identity
information)
can
be
experimentally
manipulated
by
altering
the
speed

of
object
motion
when
the
object
is
being
encoded
in
memory.
The
slower
an
object
moves

during
encoding,
the
more
identity
information
(and
less
viewpoint-‐specific
information)

becomes
encoded
in
the
chick’s
representation
of
the
object.
While
previous
research
has

begun
to
reveal
the
developmental
origins
of
object
recognition,
the
state
of
invariant
face

recognition
at
the
onset
of
vision
remains
unknown.
Can
newborn
chicks
build
view-‐

102

invariant
face
representations,
and
are
those
representations
subject
to
the
same

constraints
as
view-‐invariant
representations
of
objects?

It
has
previously
been
shown
that
newborn
chicks
are
able
to
recognize
human

faces
shown
from
familiar
viewing
angles
(Wood
&
Wood,
2015b).
Thus,
chicks
are
able
to

discriminate
the
first
face
seen
in
life
from
some
other
faces.
However,
discriminating

familiar
animations
of
faces
could
be
accomplished
by
using
a
simple
pattern-‐matching

strategy
rather
than
building
a
view-‐invariant
representation
of
faces.
Faces
pose
a

particularly
complex
case
of
invariant
recognition
because
faces
must
be
recognized
at
the

subordinate,
or
individual,
level
while
most
objects
are
generally
recognized
at
the
“basic”

level
(Damasio,
Damasio,
&
Vanhoesen,
1982;
Diamond
&
Carey,
1986;
Tarr
&
Gauthier,

2000).
Faces
are
a
prototypical
example
of
subordinate-‐level
recognition
because
all

individual
faces
share
the
same
general
configuration
(Carey,
1992;
Leibo,
Mutch,
&
Poggio,

2011).
Thus,
face
recognition
is
an
especially
difficult
task,
requiring
more
fine-‐tuned

recognition
than
basic
object
recognition.

The
Present
Experiments

Our
study
is
divided
into
two
parts.
First,
we
tested
whether
newborn
chicks
can

build
view-‐invariant
representations
of
the
first
face
seen
in
life
(Experiments
1
&
2).

Second,
we
examined
whether
the
development
of
this
ability
is
subject
to
a
slowness

constraint.
In
particular,
we
tested
whether
it
is
possible
to
systematically
manipulate
the

information
content
of
a
newborn
chick’s
first
face
representation
by
changing
the
speed
of

object
motion
during
encoding
(Experiments
3
&
4).

103

In
Experiment
1,
we
imprinted
newborn
chicks
to
a
single
human
face
rotating
20

degrees
and
tested
the
chicks’
ability
to
recognize
the
imprinted
face
across
20
viewpoints

(19
novel,
1
familiar).
The
chicks
were
able
to
distinguish
the
imprinted
face
from
the
two

unfamiliar
test
faces
across
novel
viewpoints.
In
Experiment
2,
we
controlled
for
possible

differences
in
color
across
the
imprinted
and
unfamiliar
faces
by
converting
all
of
the
faces

to
the
same
shade
of
red.
Despite
removing
a
significant
amount
of
information
about
the

faces,
the
chicks
were
still
able
to
recognize
their
imprinted
face
across
the
viewpoint

changes.
In
Experiment
3,
we
tested
whether
the
speed
of
face
motion
during
encoding

affects
the
face
representation
built
by
the
subject.
Specifically,
newborn
chicks
were

imprinted
to
a
single
face
that
rotated
at
one
of
three
possible
speeds—fast,
medium,
or

slow.
Akin
to
the
development
of
object
recognition,
we
found
a
trade-‐off
between
the

amount
of
identity
information
and
viewpoint-‐specific
information
in
the
chicks’
face

representations.
The
amount
of
identity
and
viewpoint-‐specific
information
in
the
chicks’

face
representations
was
directly
related
to
the
speed
of
the
face
during
encoding.
Finally,

in
Experiment
4
we
confirmed
that
chicks
can
extract
identity
information
from
quickly

moving
faces,
provided
that
the
face
was
moving
slowly
when
being
encoded
into
memory.

Thus,
identity
information
is
available
when
a
face
moves
quickly
or
slowly,
but
more

identity
information
is
encoded
when
a
face
moves
slowly.
Overall,
our
findings
indicate

that
invariant
face
recognition
and
invariant
object
recognition
develop
from
some

common
machinery
and
are
subject
to
the
same
developmental
constraint.

104

Experiment
1

Methods

Subjects

Ten
Rhode
Island
Red
chicks
of
unknown
sex
were
tested
in
Experiment
1.
No

subjects
were
excluded
from
the
analyses.
The
eggs
were
obtained
from
a
local
distributer

and
incubated
in
darkness
in
an
OVA-‐Easy
incubator
(Brinsea
Products
Inc.,
Titusville,
FL).

The
temperature
was
maintained
at
99.6°F
and
the
humidity
was
maintained
at
45%
for

the
first
19
days
of
incubation.
On
day
19,
we
increased
the
humidity
to
60%.
On
day
1
of

life,
the
subjects
were
moved
from
the
incubator
room
to
the
controlled-‐rearing
chambers

in
darkness
with
the
aid
of
night
vision
goggles.
All
care
of
the
subjects
was
also
done
in

darkness
with
night
vision
goggles.
Each
chick
was
housed
singly
in
its
own
chamber.
All
of

the
experiments
presented
here
were
approved
by
the
University
of
Southern
California

Institutional
Animal
Care
and
Use
Committee.

Controlled-‐Rearing
Chambers

Subjects
were
raised
from
birth
for
two
weeks
within
controlled-‐rearing
chambers

(66
cm
length
x
42
cm
width
x
69
cm
height).
The
chambers
were
constructed
from
white,

high-‐density
plastic.
For
a
picture
of
the
chambers,
see
Figure
1
in
Wood
(2013).
Face

stimuli
were
presented
to
the
subjects
by
projecting
animated
videos
onto
two
display

walls
(19”
liquid
crystal
display
monitors
with
1440
x
900
pixel
resolution)
situated
on

opposite
sides
of
the
chamber.
The
chambers
contained
no
rigid,
bounded
objects
other

than
the
virtual
face
presented
on
the
display
walls.
All
care
of
the
subjects
was
performed

in
darkness
with
the
aid
of
night
vision
goggles.
Food
and
water
were
provided
ad
libitum

105

in
transparent,
rectangular
holes
in
the
ground
(66
cm
length
x
2.5
cm
width
x
2.7
cm

height).
We
used
grain
as
food
because
grain
does
not
behave
like
an
object
(i.e.,
grain
does

not
maintain
a
rigid,
bounded
shape).
The
floors
were
black
wire
mesh
supported
over
a

black
surface
by
thin,
transparent
beams.

The
chicks’
behavior
was
tracked
by
micro-‐cameras
(1.5
cm
diameter)
embedded
in

the
ceilings
of
the
chambers
and
Ethovision
XT
software
(Noldus
Information
Technology).

This
software
calculated
the
amount
of
time
each
subject
spent
within
zones
(22
cm
×
42

cm)
next
to
the
left
and
right
display
walls.

Input
Phase

In
the
input
phase
(the
first
week
of
life),
subjects
were
raised
in
a
visual

environment
that
contained
a
single
virtual
face
(ear-‐to-‐ear
width
=
6.2-‐6.5
cm;
height
=
10

cm;
distance
above
flooring
unit
=
1
cm).
Six
of
the
chicks
were
imprinted
to
Face
A,
and

four
were
imprinted
to
Face
B
(see
Figure
24).
The
virtual
face
moved
continuously,

rotating
through
a
20°
viewing
range
about
a
vertical
axis
passing
through
its
centroid.
The

individual
frames
of
face
movement
were
created
using
FaceGen
software
(Singular

Inversion,
Inc.)
and
concatenated
into
an
animation
using
QTCoffee
(3AM
Coffee
Software).

The
face
was
shown
on
a
uniform
white
background.
During
the
input
phase,
the
imprinted

face
switched
display
walls
every
2
hours,
following
a
one-‐minute
period
of
darkness

(Figure
25).
The
face
appeared
for
an
equal
amount
of
time
on
each
display
wall.

106

Test
Phase

In
the
test
phase
(the
second
week
of
life),
we
tested
whether
the
chicks
were
able

to
build
a
view-‐invariant
representation
of
their
imprinted
face.
To
do
so,
we
used
a
two-‐
alternative
forced
choice
test.
On
each
test
trial,
the
imprinted
face
was
projected
on
one

display
wall
and
an
unfamiliar
face
was
projected
on
the
other
display
wall.
The
familiar

face
was
presented
rotating
through
a
20°
viewpoint
range
from
one
of
20
possible
viewing

angles
(the
imprinted
viewing
angle
plus
19
novel
viewing
angles).
The
unfamiliar
faces

Figure
24.
Images
from
the
Face
A
animation
and
Face
B
animation
shown
during
the

input
phase
in
Experiment
1
(top)
and
Experiment
2
(bottom).
The
faces
rotated

smoothly
and
continuously
through
a
20
degree
viewpoint
range.
Each
chick
saw
a

single
face
(either
Face
A
or
Face
B)
during
the
input
phase.

107

had
the
same
size,
motion
speed,
and
viewpoint
range
as
the
imprinted
face
from
the
input

phase.
Consequently,
on
most
of
the
test
trials,
the
unfamiliar
face
was
more
similar
to
the

imprinting
stimulus
than
the
imprinted
face
was
to
the
imprinting
stimulus
(from
a
pixel-‐
wise
perspective).
To
recognize
their
imprinted
face,
the
chicks
therefore
needed
to

generalize
across
large
changes
in
the
face’s
appearance.
If
the
chicks
could
recognize
their

imprinted
face,
then
they
should
spend
a
greater
proportion
of
time
in
proximity
to
the

imprinted
face
compared
to
the
unfamiliar
face
during
these
test
trials.

Figure
25.
A
schematic
of
the
presentation
of
the
virtual
faces
during
the
input
phase

(top)
and
test
phase
(bottom).
During
the
input
phase
the
chicks
were
raised
with
a

single
face
rotating
20
degrees
back
and
forth
about
a
frontoparallel
axis
through
the

face’s
centroid.
The
test
phase
consisted
of
alternating
test
trials
and
rest
periods.

During
each
test
trial,
the
imprinted
face
was
shown
on
one
display
wall
and
an

unfamiliar
face
was
shown
on
the
opposite
display
wall.
The
imprinted
face
was
shown

from
a
variety
of
novel
viewpoints
across
the
test
trials,
while
the
unfamiliar
face
was

always
shown
from
the
same
viewpoint
range
as
the
imprinted
face
during
the
rest

periods
(to
maximize
the
image-‐level
similarity
between
the
unfamiliar
face
and

imprinted
stimulus).

108

The
test
trials
were
12
minutes
in
duration
followed
by
24-‐minute
rest
trials
(Figure

25).
During
the
rest
trials,
the
imprinted
face
appeared
on
one
display
wall
and
a
white

screen
appeared
on
the
other
display
wall.
Each
subject
received
40
test
trials
per
day.

Figure
26.
Results
from
Experiments
1
&
2.
The
circle
charts
(top)
depict
the
test

viewpoints
of
the
imprinted
face.
Each
test
viewpoint
is
color-‐coded
based
on
the

average
performance.
White
represents
chance
performance
(50%)
and
green

represents
ceiling
performance
(defined
as
the
average
preference
for
the
imprinted

face
over
the
blank
display
wall
during
rest
trials
for
that
experiment).
The
bar
graphs

(bottom)
show
performance
by
degree
of
viewpoint
change
(i.e.,
the
degrees
of
change

between
the
middle
frame
of
the
imprinted
face
animation
and
the
middle
frame
of
the

test
face
animation).
Chicks’
performance
was
above
chance
for
all
three
of
the

viewpoint
change
categories
(0°,
25°,
and
50°)
in
Experiments
1
&
2.

109

Results

Results
are
shown
in
Figure
26.
To
measure
the
chicks’
performance,
we
computed

the
percent
of
time
the
chicks
spent
with
the
imprinted
face
compared
to
the
unfamiliar

face
on
the
trials
in
which
the
imprinted
face
switched
display
walls
from
the
preceding

rest
period
and
the
trials
in
which
the
imprinted
face
did
not
switch
display
walls
from
the

preceding
rest
period.
Then
we
computed
the
average
of
these
two
values
to
obtain
a
single

recognition
performance
score
for
each
chick.
Across
all
of
the
test
trials,
the
newborn

chicks
spent
significantly
more
time
with
their
imprinted
face
than
the
unfamiliar
face
(M
=

67%,
SD
=
5%;
one-‐sample
t-‐test,
t(9)
=
10.76,
p
=
0.000002,
d

=
3.40).
Mean
performance

was
the
same
after
removing
test
trials
in
which
the
imprinted
face
was
shown
from
the

familiar
viewpoint
range
and
remained
well-‐above
chance
levels
(M
=
67%,
SD
=
5%;
one-‐
sample
t-‐test,
t(9)
=
11.13,
p
=
0.000001,
d
=
3.52).

To
examine
whether
chicks
showed
impaired
performance
for
larger
viewpoint

changes,
we
classified
each
of
the
test
trial
viewpoint
ranges
into
three
categories:
0°

viewpoint
change
(the
middle
frame
of
the
viewpoint
range
in
the
test
animation
was
the

same
as
the
middle
frame
of
the
viewpoint
range
in
the
imprinted
animation),
±25°

viewpoint
change
(the
middle
frame
of
the
viewpoint
range
in
the
test
animation
was
±25°

from
the
middle
frame
of
the
viewpoint
range
in
the
imprinted
animation),
and
±50°

viewpoint
change
(the
middle
frame
of
the
viewpoint
range
in
the
test
animation
was
±50°

from
the
middle
frame
of
the
viewpoint
range
in
the
imprinted
animation).
We
then

performed
a
repeated-‐measures
ANOVA
with
the
within-‐subjects
main
effect
of
viewpoint

change.
We
found
a
significant
effect
of
viewpoint
change
(F(2,
18)
=
6.89,
p
=
.006,
η
2

=

.43).
Post-‐hoc
paired-‐sample
t-‐tests
revealed
that
performance
in
the
0°
test
trials
was

110

significantly
higher
than
both
the
±25°
test
trials
(t(9)
=
2.66,
p
=
.026,
d
=
.84)
and
the
±50°

test
trials
(t(9)
=
3.26,
p
=
.010,
d
=
1.03).
However,
it
is
important
to
note
that
performance

was
still
well
above
chance
levels
in
all
of
these
viewpoint
ranges
(one-‐sample
t-‐tests,
0°:

t(9)
=
9.76,
p
=
.000004,
d
=
3.09;
±25°:
t(9)
=
10.35,
p
=
.000003,
d
=
3.27;
±50°:
t(9)
=
8.20,

p
=
.00002,
d
=
2.59).

To
test
whether
performance
varied
as
a
function
of
the
day
of
testing,
we

performed
a
repeated-‐measures
ANOVA
with
the
within-‐subjects
main
effect
of
test
day.

The
ANOVA
did
not
reveal
a
significant
main
effect
of
test
day
(F(6,
54)
=
.83,
p
=
.55).

Moreover,
performance
was
above
chance
levels
on
all
test
days
(one-‐sample
t-‐tests,
Holm-‐
Bonferroni
corrected
for
multiple
comparisons,
all
Ps
<
.001).
Therefore,
performance
did

not
vary
significantly
by
test
day.

Discussion

The
results
of
Experiment
1
suggest
that
newborn
chicks
are
able
to
build
a
view-‐
invariant
representation
of
the
first
face
seen
in
life.
However,
the
face
stimuli
used
in

Experiment
1
had
slightly
different
colors.
Thus,
based
on
Experiment
1,
we
could
not

exclude
the
possibility
that
the
chicks
were
relying
on
color
alone
to
recognize
their

imprinted
face.
To
control
for
this
possibility,
in
Experiment
2,
we
converted
all
of
the
faces

to
red-‐scale
to
remove
any
hue-‐based
identity
cues.

111

Experiment
2

Methods

The
methods
in
Experiment
2
were
identical
to
Experiment
1,
with
the
following

two
exceptions.
First,
a
new
group
of
12
chicks
were
tested.
Second,
the
face
stimuli
were

converted
to
red-‐scale
to
control
for
color
differences
(see
Figure
24).

Results

The
results
are
shown
in
Figure
26.
Newborn
chicks
were
able
to
recognize
their

imprinted
face
across
the
test
trials,
despite
the
removal
of
any
hue
differences
(M
=
56%,

SD
=
6%;
one-‐sample
t-‐test,
t(11)
=
3.83,
p
=
.003,
d
=
1.11).
Mean
performance
remained

above
chance
levels
even
after
removing
the
test
trials
in
which
the
imprinted
face
was

shown
from
the
familiar
(imprinted)
viewpoint
range
(M
=
57%,
SD
=
6%;
one-‐sample
t-‐
test,
t(11)
=
3.95,
p
=
.002,
d
=
1.14).

As
in
Experiment
1,
we
tested
whether
chicks
showed
impaired
performance
for

larger
viewpoint
changes.
A
repeated-‐measures
ANOVA
with
the
within-‐subjects
main

effect
of
viewpoint
range
did
not
reveal
a
main
effect
of
viewpoint
angle
(F(2,
22)
=
.94,
p
=

.41).
Performance
was
above
chance
on
all
viewpoint
range
groups
(one-‐sample
t-‐tests,
0°:

t(11)
=
3.07,
p
=
.01,
d
=
.89;
±25°:
t(11)
=
3.72,
p
=
.003,
d
=
1.07;
±50°:
t(11)
=
3.57,
p
=

.004,
d
=
1.03).

We
also
tested
whether
performance
varied
as
a
function
of
the
day
of
testing.
A

repeated-‐measures
ANOVA
with
the
within-‐subjects
main
effect
of
test
day
revealed
a
non-‐
significant
(but
trending)
main
effect
of
test
day
(F(6,
66)
=
2.12,
p
=
.06,
η
2

=
.16).

Performance
was
above
chance
levels
on
all
test
days
(one-‐sample
t-‐tests,
all
Ps
<
.05);

112

however,
only
performance
on
day
1
and
day
3
were
above
chance
levels
after
Holm-‐
Bonferroni
correction.

Finally,
we
tested
whether
performance
was
hindered
by
the
removal
of
color
cues

of
facial
identity.
Independent
samples
t-‐tests
showed
that
performance
was
significantly

higher
in
Experiment
1
(i.e.,
with
color
cues)
than
in
Experiment
2
(i.e.,
without
color
cues),

(t(20)
=
4.60,
p
=
.0002).
Thus,
while
color
cues
are
not
necessary
for
chicks
to
perform

invariant
face
recognition,
color
cues
can
improve
performance
significantly.

Discussion

Performance
in
Experiment
2
was
significantly
lower
than
Experiment
1;
however,

performance
in
both
experiments
was
above
chance
levels.
The
decrease
in
performance

from
Experiment
1
to
Experiment
2
is
unsurprising
given
that
color
is
an
important
cue
for

face
recognition
even
in
mature
visual
systems
(Farah
et
al.,
1998;
Hill
et
al.,
1995;
Said
&

Todorov,
2011).
Taken
together,
Experiments
1
and
2
provide
evidence
that
newborn

chicks
are
able
to
build
a
view-‐invariant
representation
of
the
first
face
seen
in
life.
In

Experiments
3
&
4,
we
tested
whether
the
development
of
invariant
face
representations
is

hard-‐wired
or
dependent
on
the
chicks’
visual
experiences
with
faces.

113

Experiment
3

Methods

Subjects

Thirty-‐five
Rhode
Island
Red
chicks
of
unknown
sex
were
tested
in
Experiment
3

(11
to
12
subjects
per
speed
condition).
No
subjects
were
excluded
from
the
analyses.
The

incubation
procedure
and
the
test
chambers
were
identical
to
Experiments
1
and
2.

Figure
27.
A
schematic
of
the
presentation
of
the
virtual
faces
during
the
input
phase

(top)
and
test
phase
(bottom)
for
Experiment
3.
During
the
input
phase
the
chicks
were

raised
with
a
single
face
rotating
120
degrees
back
and
forth
about
a
frontoparallel
axis

through
the
face’s
centroid.
The
test
phase
consisted
of
alternating
test
trials
and
rest

periods.
During
each
Viewpoint
test
trial,
the
imprinted
face
was
shown
from
the

familiar
viewpoint
range
on
one
display
wall
and
from
an
unfamiliar
viewpoint
range
on

the
opposite
display
wall.
During
each
Identity
test
trial,
the
imprinted
face
was
shown

from
an
unfamiliar
viewpoint
range
on
one
display
wall
and
the
opposite
display

showed
an
unfamiliar
face
from
the
same
viewpoint
range
as
the
imprinted
face
during

the
rest
periods.

114

Input
Phase

In
the
input
phase
(the
first
week
of
life),
subjects
were
raised
in
a
visual

environment
that
contained
a
single
virtual
face.
The
virtual
face
moved
continuously,

rotating
through
a
120°
viewing
range
about
a
vertical
axis
passing
through
its
centroid.

Each
chick
was
randomly
assigned
to
either
the
Slow,
Medium
or
Fast
Condition.
In
the

Slow
Condition,
the
virtual
face
rotated
back
and
forth
120°
in
20s.
In
the
Medium

Condition,
the
virtual
face
rotated
back
and
forth
120°
in
5s.
In
the
Fast
Condition,
the

virtual
face
rotated
back
and
forth
120°
in
1s.

The
same
two
imprinted
faces
were
used
as
in
Experiment
1,
with
half
of
the
chicks

imprinted
to
Face
A
and
the
other
half
imprinted
to
Face
B.
All
faces
were
shown
on
a

uniform
white
background.
During
the
input
phase,
the
imprinted
face
switched
display

walls
every
2
hours,
following
a
one-‐minute
period
of
darkness
(Figure
27).
The
face

appeared
for
an
equal
amount
of
time
on
each
display
wall.

Test
Phase

During
the
test
phase
(the
second
week
of
life),
we
probed
the
informational
content

of
the
face
representation
built
by
each
subject,
by
using
an
automated
two-‐alternative

forced
choice
testing
procedure.
Subjects
were
expected
to
spend
a
greater
proportion
of

time
in
proximity
to
the
object
that
they
perceived
to
be
their
imprinted
object.
The
test

trials
lasted
20
minutes
and
were
separated
from
one
another
by
40-‐minute
rest
periods

(see
Figure
27).
During
the
rest
periods,
the
imprinted
face
appeared
on
one
display
wall

and
a
white
screen
appeared
on
the
other
display
wall.
Each
subject
received
up
to
24
test

115

trials
per
day.
The
conditions
were
presented
in
randomized
blocks
throughout
the
test

phase.

Subjects
were
presented
with
two
types
of
test
trials
(see
Figure
28
for

illustrations):

1. Identity
Trials:
The
imprinted
face
was
paired
with
an
unfamiliar
face
that
had
a

similar
size
and
motion
speed
as
the
imprinted
face.
In
each
Identity
Trial,
the

imprinted
face
rotated
120°
degrees
through
a
novel
axis
(an
axis
tilted
45°

Figure
28.
A
visualization
of
the
imprinted
animations
and
the
test
trial
stimuli.
The
top

images
show
frames
from
the
imprinting
animation
for
the
female
face
(A)
and
male
face

(B).
During
the
test
phase,
subjects
were
given
Viewpoint
Trials
and
Identity
Trials.
In

the
Viewpoint
Trials
(blue
boxes),
one
display
wall
showed
the
imprinted
face
from
the

familiar
viewpoint
and
the
other
display
wall
showed
the
imprinted
face
from
an

unfamiliar
viewpoint.
In
the
Identity
Trials
(green
boxes),
one
display
wall
showed
the

imprinted
face
from
an
unfamiliar
viewpoint
and
the
other
display
wall
showed
an

unfamiliar
face
from
the
familiar
viewpoint.
A
fully
invariant
face
representation
should

be
sensitive
to
identity
cues
but
not
viewpoint-‐specific
cues.

116

diagonally
or
a
perpendicular
axis),
and
the
unfamiliar
face
rotated
identically
to

the
imprinting
animation.
In
each
Identity
Trial,
the
unfamiliar
face
was
either

the
face
that
the
other
half
of
the
chicks
were
imprinted
to
(e.g.,
Face
B
if
a

subject
was
imprinted
to
Face
A)
or
a
novel
elderly
face.

2. Viewpoint
Trials:
One
display
wall
showed
the
imprinted
face
rotating
around

the
familiar
axis,
while
the
other
display
wall
showed
the
imprinted
face
rotating

around
a
novel
axis.
The
two
novel
axes
were
an
axis
tilted
45°
diagonally
and
a

perpendicular
axis.

Results

Our
main
research
hypothesis
was
that
the
speed
of
face
motion
would
affect
the

informational
content
(identity
versus
viewpoint-‐specific
information)
of
the
chicks’
face

representations.
To
test
this
hypothesis,
we
performed
a
repeated-‐measures
ANOVA
with

the
within-‐subjects
main
effect
of
Trial
Type
(Identity
vs.
Viewpoint
trials)
and
the

between-‐subjects
main
effect
of
Face
Speed
(Slow,
Medium,
or
Fast).
The
ANOVA
revealed
a

significant
main
effect
of
Trial
Type
(F(1,
32)
=
24.85,
p
=
.00002,
η
2

=
.44)
and
a
significant

interaction
between
Trial
Type
and
Face
Speed
(F(2,
32)
=
22.72,
p
=
.00000007,
η
2

=
.59).

The
main
effect
of
Face
Speed
was
not
significant.
The
significant
interaction
was
driven
by

higher
performance
in
Identity
Trials
for
slower
speeds
of
face
motion
and
higher

performance
in
Viewpoint
Trials
for
faster
speeds
of
face
motion
(see
Figure
29).

Additionally,
we
performed
planned
follow-‐up
analyses
to
determine
the
conditions

in
which
the
chicks
performed
above
chance
levels.
We
found
that
chicks
performed

significantly
above
chance
levels
on
Identity
Trials
in
the
Slow
Condition
(t(10)
=
13.56,
p
=

117

.0000001,
d
=
4.09),
the
Medium
Condition
(t(11)
=
7.15,
p
=
.00002,
d
=
2.06),
and
the
Fast

Condition
(t(11)
=
3.48,
p
=
.005,
d
=
1.01).
Conversely,
we
found
that
the
chicks
performed

significantly
above
chance
levels
on
Viewpoint
Trials
in
the
Fast
Condition
(t(11)
=
11.19,
p

=
.0000002,
d
=
3.23)
and
the
Medium
Condition
(t(11)
=
3.58,
p
=
.004,
d
=
1.03),
but
not
in

the
Slow
Condition
(t(10)
=
1.66,
p
=
.13,
d
=
-‐.50).

Figure
29.
Results
of
Experiment
3
(A)
and
Experiment
4
(B).
In
Experiment
3,
chicks

were
imprinted
to
a
face
that
rotated
at
a
slow,
medium,
or
fast
speed,
and
the
speed
of

face
motion
during
the
test
trials
matched
the
speed
of
face
motion
during
imprinting.

Performance
in
the
test
trials
was
directly
related
to
the
speed
of
the
imprinted
face.
The

slower
the
imprinted
face
moved,
the
better
the
performance
in
the
Identity
Trials
and

the
worse
the
performance
in
the
Viewpoint
Trials.
In
Experiment
4,
all
chicks
were

imprinted
to
a
slowly
moving
face,
but
the
speed
of
the
face
varied
across
test
trials.

When
the
face
moved
slowly
during
the
input
phase,
chicks
performed
similarly
to
the

chicks
in
the
slow
condition
of
Experiment
3,
regardless
of
the
speed
of
the
object
during

the
test
trial.

118

To
test
whether
performance
varied
as
a
function
of
the
test
day,
we
computed
a

repeated-‐measures
ANOVA
with
the
within-‐in
subjects
main
effects
of
Trial
Type
and
Test

Day
and
the
between-‐subjects
main
effect
of
Face
Speed.
As
above,
the
ANOVA
revealed
a

significant
main
effect
of
Trial
Type
(F(1,
31)
=
22.28,
p
=
.00005,
η
2

=
.42)
and
a
significant

interaction
between
Trial
Type
and
Face
Speed
(F(2,
31)
=
19.23,
p
=
.000004,
η
2

=
.55).
The

ANOVA
also
showed
a
significant
main
effect
of
Test
Day
(F(6,
186)
=
2.22,
p
=
.04,
η
2

=
.07)

and
a
significant
interaction
of
Test
Day
and
Face
Speed
(F(12,
186)
=
2.39,
p
=
.007,
η
2

=

.13).
The
main
effect
of
Face
Speed,
the
interaction
of
Trial
Type
and
Test
Day,
and
the

interaction
of
Trial
Type,
Test
Day,
and
Face
Speed
were
not
significant
(Ps
>
.05).

However,
overall
performance
was
above
chance
on
the
first
day
of
testing
for
all
imprinted

face
speeds
(two-‐tailed
one
sample
t-‐tests,
all
Ps
<
.01).

Discussion

Experiment
3
tested
whether
the
speed
of
face
movement
affected
newborn
chicks’

ability
to
perform
view-‐invariant
face
recognition.
Two
main
findings
emerged.
First,
we

found
a
trade-‐off
between
the
amount
of
identity
and
view-‐specific
face
information

encoded
by
the
chicks.
Second,
we
found
a
relationship
between
the
speed
of
face
motion

and
the
information
encoded
in
the
chicks’
representations.
When
the
faces
moved
slowly,

newborn
chicks
built
view-‐invariant
face
representations
that
were
highly
sensitive
to

identity
information
and
tolerant
to
changes
in
viewpoint.
Conversely,
when
the
faces

moved
quickly,
newborn
chicks
built
face
representations
that
were
less
sensitive
to

identity
information
and
more
selective
for
familiar
viewpoints.
Thus,
there
is
a
slowness

constraint
on
the
development
of
invariant
face
recognition.

119

Because
the
faces
in
Experiment
3
moved
at
the
same
speed
during
the
input
and

test
phases,
it
is
possible
that
the
chicks’
impairment
at
recognizing
quickly
moving
faces
is

due
to
is
due
to
limitations
in
their
ability
to
attend
to
or
perceive
quickly
moving
faces.
We

have
previously
found
that
this
explanation
cannot
account
for
the
slowness
constraint
in

the
domain
of
object
recognition
(Wood
&
Wood,
2016a).
In
particular,
newborn
chicks
can

successfully
recognize
quickly
moving
objects
provided
that
the
object
moved
slowly
when

being
encoded
into
memory.
In
Experiment
4,
we
tested
whether
a
similar
pattern
occurs

in
the
development
of
face
recognition.

Experiment
4

Methods

The
methods
for
Experiment
4
were
identical
to
Experiment
3
with
three

exceptions.
First,
a
new
group
of
12
chicks
were
tested.
Second,
all
of
the
chicks
were

imprinted
to
the
slowly
moving
face
during
the
input
phase
(i.e.,
the
input
phase
for
all

chicks
in
Experiment
4
was
identical
to
the
Slow
Condition
input
phase
in
Experiment
3).

Third,
during
the
test
phase,
the
chicks
were
tested
with
faces
that
moved
at
the
fast,

medium,
and
slow
speeds
from
Experiment
3.
Thus,
in
Experiment
4,
each
chick
received
6

types
of
trials:
Slow
Identity,
Slow
Viewpoint,
Medium
Identity,
Medium
Viewpoint,
Fast

Identity,
and
Fast
Viewpoint
(see
Figure
29).
If
fast
motion
impairs
newborn
chicks’
ability

to
perceive
faces,
then
performance
in
the
Identity
Trials
should
be
lower
during
the
Fast

test
trials
than
the
Slow
test
trials.
Conversely,
if
fast
motion
primarily
affects
face

encoding,
then
performance
should
be
equally
high
in
the
Identity
Trials
across
all
of
the

test
speeds
(because
the
face
moved
slowly
when
being
encoded
into
memory).

120

Results

Results
are
shown
in
Figure
29.
The
main
analysis
of
interest
is
whether

performance
in
the
Identity
Trials
varied
as
a
function
of
the
face
speed
during
testing.
To

assess
this
question,
we
performed
a
repeated
measures
ANOVA
with
the
within-‐subjects

factors
of
Face
Speed
(Slow,
Medium,
Fast)
and
Trial
Type
(Identity
and
Viewpoint).
The

main
effect
of
Trial
Type
was
significant
(F(1,
11)
=
90.85,
p
=
.000001,
η
2

=
.89)
as
was
the

main
effect
of
Face
Speed
(F(2,
22)
=
4.07,
p
=
.03,
η
2

=
.27).
The
interaction
of
Trial
Type

and
Face
Speed
was
not
significant
(F(2,
22)
=
2.60,
p
=
.10,
η
2

=
.19).

For
the
main
effect
of
Trial
Type,
performance
was
significantly
higher
on
the

Identity
Trials
than
the
Viewpoint
Trials
(paired
t-‐test,
t(11)
=
9.27,
p
=
.000002,
d
=
2.68).

Performance
was
significantly
higher
on
the
Identity
Trials
than
Viewpoint
Trials
at
every

testing
speed
(paired
t-‐tests,
all
Ps
Holm-‐Bonferroni
corrected;
Slow:
t(11)
=
7.67,
p
=

.00002,
d
=
2.21;
Medium:
t(11)
=
8.02,
p
=
.00002,
d
=
2.31;
Fast:
t(11)
=
4.14,
p
=
.002,
d
=

1.20).
These
results
are
consistent
with
performance
in
the
Slow
Condition
of
Experiment

3.

For
the
main
effect
of
Face
Speed,
was
performance
significantly
hindered
by
fast

moving
faces?
Paired
t-‐tests
revealed
that
overall
performance
was
not
lower
for
faster

face
speeds.
In
fact,
overall
performance
was
significantly
higher
for
the
Fast
trials
than

Slow
trials
(t(11)
=
2.91,
p
=
.01,
d
=
.84;
Fast
vs.
Medium
and
Slow
vs.
Medium
were
not

significant,
p
>
.05).

To
directly
test
whether
performance
in
the
Identity
Trials
varied
according
to
the

Face
Speed
during
testing,
we
conducted
a
repeated
measures
ANOVA
on
the
Identity

121

Trials
alone
(no
Viewpoint
Trials)
with
the
within-‐subjects
factor
of
Face
Speed.
The

ANOVA
found
no
significant
differences
in
performance
across
the
three
face
speeds
in
the

Identity
Trials
(F(2,
22)
=
2.14,
p
=
.14,
η
2

=
.16).

Discussion

When
chicks
were
raised
with
a
slowly
moving
face,
they
built
face
representations

that
were
highly
sensitive
to
identity
information
and
insensitive
to
viewpoint-‐specific

information,
regardless
of
the
speed
of
face
motion
during
testing.
Therefore,
the
results
of

Experiment
3
cannot
be
explained
by
a
limitation
in
chicks’
ability
to
perceive
or
attend
to

quickly
moving
faces.
Overall,
the
results
of
Experiments
3
and
4
demonstrate
that
it
is

possible
to
manipulate
a
newborn’s
first
face
representation
by
changing
the
rotation

speed
of
the
face
during
encoding.

General
Discussion

This
study
examined
whether
a
newborn
animal—the
domestic
chick—is
capable
of

building
an
invariant
representation
of
the
first
face
seen
in
life.
We
found
that
newborn

chicks
can
build
view-‐invariant
face
representations
at
the
onset
of
vision,
successfully

recognizing
familiar
faces
across
a
wide
range
of
novel
viewpoints.
Additionally,
we
found

that
the
development
of
this
ability
is
subject
to
a
slowness
constraint.
When
newborn

chicks
were
raised
with
a
slowly
moving
face,
they
built
invariant
face
representations
that

were
sensitive
to
identity
information
and
tolerant
to
viewpoint
changes.
Conversely,
when

newborn
chicks
were
raised
with
a
quickly
moving
face,
they
built
inaccurate
face

representations
that
were
sensitive
to
viewpoint
information
but
not
identity
information.

122

Thus,
it
is
possible
to
systematically
manipulate
the
information
content
of
a
newborn

chick’s
first
face
representation
simply
by
changing
the
speed
of
face
motion
during

encoding.

The
present
study
builds
on
prior
work
showing
that
newborn
chicks
are
sensitive

to
face-‐like
configurations
of
features
(Rosa-‐Salva
et
al.,
2010)
and
can
recognize
human

faces
at
the
onset
of
vision
(Wood
&
Wood,
2015b).
Our
work
extends
these
findings
by

demonstrating
that
newborn
chicks
are
able
to
recognize
their
imprinted
face
over

significant
variation
in
the
retinal
image
produced
by
the
face.
These
results
also
extend

prior
work
by
showing
that
the
development
of
face
recognition,
like
the
development
of

object
recognition,
is
subject
to
a
slowness
constraint
(Wood
&
Wood,
2016a).
In
order
to

build
invariant
representations,
newborn
brains
require
visual
experience
with
slowly

moving
objects
and
faces.
This
finding
indicates
that
object
and
face
recognition
emerge

from
some
similar
computational
operations.

In
particular,
these
findings
are
compatible
with
unsupervised
temporal
learning

(UTL)
models
of
vision
(Li
&
DiCarlo,
2008).
According
to
UTL
models,
the
brain
constructs

invariant
object
representations
by
encoding
the
spatiotemporal
statistics
produced
by

consecutive
retinal
images
of
an
object
(Masquelier
&
Thorpe,
2007;
Wallis
&
Rolls,
1997;

Wiskott
&
Sejnowski,
2002;
Wyss,
Konig,
&
Verschure,
2006).
Since
an
object’s
identity
is

temporally
stable,
different
retinal
images
of
the
same
object
tend
to
be
contiguous
over

time.
UTL
models
capitalize
on
this
principle
by
associating
inputs
that
occur
closely

together
in
time.
Thus,
slower
movement
may
produce
more
precise
features
(as
fewer

photoreceptor
cells
are
activated
in
a
single
time
window),
while
faster
movement
may

123

produce
features
that
are
more
blurred
(as
more
photoreceptor
cells
are
activated
in
a

single
time
window).

To
what
extent
can
our
results
illuminate
the
development
of
invariant
face

recognition
in
humans?
Our
study
focused
on
chicks
as
an
animal
model
for
invariant

recognition
because
chicks
can
be
raised
from
birth
in
controlled-‐rearing
chambers

without
parental
care,
allowing
full
control
of
their
visual
development.
Importantly,

mammalian
and
avian
brains
share
basic
neuronal
types
and
anatomical
connectivity

patterns
(Dugas-‐Ford
et
al.,
2012;
Shanahan
et
al.,
2013;
Wang
et
al.,
2010).
Chickens
also

share
a
number
of
cognitive
abilities
with
humans
(Vallortigara,
2012)
including

recognizing
partly
occluded
objects
(Regolin
&
Vallortigara,
1995),
tracking
hidden
objects

(Prasad,
Wood,
&
Wood,
in
prep;
Vallortigara,
Regolin,
Rigoni,
&
Zanforlin,
1998),
and

reasoning
about
physical
interactions
between
objects
(Chiandetti
&
Vallortigara,
2011).

These
similarities
suggest
that
our
results
may
have
broader
implications
for
the

development
of
face
recognition
abilities
in
humans
as
well.

More
generally,
this
study
provides
the
first
systematic
examination
of
newborns’

ability
to
build
invariant
face
representations.
Our
findings
demonstrate
that
newborn

visual
systems
can
create
face
representations
that
are
highly
generative,
extending
beyond

raw
retinal
inputs.
However,
newborns’
development
of
invariant
face
recognition
is

subject
to
a
slowness
constraint.
Experience
viewing
slowly
moving
faces
is
critical
to
the

development
of
invariant
face
recognition.
Thus,
newborns
learn
invariant
face
recognition

through
experience
with
a
slowly
changing
visual
world.

124

Chapter
7:
Conclusion

“No
phenomenon
in
nature
is
harder
to
explain
than
the
transactions
that
are
carried
on

between
the
mind
and
the
external
world;
there
is
no
phenomenon
that
philosophical

minds
have
been
more
eager
to
dig
into
and
to
resolve.”

Thomas
Reid,
An
Inquiry
into
the
Human
Mind,
1764

“There
is
no
doubt
whatever
that
all
our
cognition
begins
with
experience;
for
how
else

should
the
cognitive
faculty
be
awakened
into
exercise
if
not
through
objects
that
stimulate

our
senses
and
in
part
themselves
produce
representations…
But
although
all
our
cognition

commences
with
experience
yet
it
does
not
on
that
account
all
arise
from
experience.
For
it

could
well
be
that
even
our
experiential
cognition
is
a
composite
of
that
which
we
receive

through
impressions
and
that
which
our
own
cognitive
faculty…
provides
out
of
itself…
It
is

therefore
at
least
a
question
requiring
closer
investigation,
and
one
not
to
be
dismissed
at

first
glance,
whether
there
is
any
cognition
independent
of
all
experience
and
even
of
all

impressions
of
the
senses.”

Immanuel
Kant,
A
Critique
of
Pure
Reason,
1787

For
centuries,
philosophers
have
debated
the
origins
of
perception
and
cognition.

The
adult
mind
instinctively
translates
a
barrage
of
sensory
input
into
meaningful

information
about
the
external
world.
To
rapidly
and
automatically
interpret
visual
input,

however,
the
brain
must
solve
a
major
computational
problem.
While
the
appearance
of

objects
on
the
retina
may
vary
radically
due
to
changes
in
viewpoint,
background,
position,

lighting,
etc.,
we
nonetheless
perceive
enduring
and
unified
objects
that
persist
through

time.
Until
recently,
it
has
been
impossible
to
investigate
how
this
ability
emerges,
due
to

the
difficulty
of
testing
newborn
subjects
experimentally.
The
goal
of
this
dissertation
was

to
use
a
recently
developed
automated
controlled-‐rearing
method
to
investigate
which

object
recognition
abilities
are
present
at
birth
and
how
those
abilities
are
shaped
by
visual

experiences.

125

Main
findings

Newborns
are
able
to
build
an
invariant
representation
of
the
first
object
seen
in
life

The
results
of
this
dissertation
provide
evidence
that
newborns
are
capable
of

forming
object
representations
that
generalize
to
novel
viewing
situations.
The
results

from
Chapter
2
demonstrate
that
newborn
chicks
can
build
a
background-‐invariant
object

representation
of
the
first
object
seen
in
life.
While
other
researchers
have
examined
visual

parsing
in
infants
and
patients
recovering
from
blindness,
this
is
the
first
experiment
to

test
background-‐invariant
recognition
at
the
onset
of
vision.
We
imprinted
newborn
chicks

to
a
single
object
rotating
on
a
single
background.
We
then
tested
whether
the
chicks
could

recognize
the
object
when
it
was
presented
from
novel
viewpoints
and
novel
backgrounds.

The
chicks
successfully
recognized
their
imprinted
object
across
novel
viewpoints
and

backgrounds,
with
no
cost
in
performance
compared
to
when
the
object
was
presented

from
the
familiar
viewpoint
on
the
familiar
background.
Additionally,
the
chicks
showed
no

evidence
of
imprinting
to
the
background.
These
results
show
that
newborn
chicks
are

capable
of
segmenting
objects
from
backgrounds
and
building
abstract
object

representations
that
generalize
across
novel
viewing
situations.

Furthermore,
the
results
from
Chapter
4
demonstrate
that
newborn
chicks
can
build

view-‐invariant
representations
from
extremely
sparse
visual
input.
In
Chapter
4,
chicks

were
imprinted
to
three
frames
of
a
single
object
during
the
input
phase.
During
the
test

phase,
the
chicks
were
tested
on
their
ability
to
recognize
the
object
across
26
novel

viewpoint
ranges.
Thus,
to
recognize
their
imprinted
object,
the
chicks
needed
to
build
a

view-‐invariant
representation
of
the
object
from
only
three
object
images.
This
is

extremely
sparse
input,
even
relative
to
the
previous
automated
controlled-‐rearing

126

experiments
discussed
in
the
Introduction.
For
example,
in
Wood
(2013)
the
imprinted

object
rotated
60°
presenting
72
unique
views
of
the
object.
Despite
receiving
only
three

views
of
the
object,
the
chicks
still
successfully
recognized
their
imprinted
object
across

novel
viewpoints.
(Although
notably,
their
representations
were
not
fully
invariant—the

chicks
were
unable
to
recognize
the
familiar
object
when
the
viewpoint
angle
produced
too

much
self-‐occlusion.)
These
results
indicate
that
newborn
neural
circuits
are
surprisingly

powerful,
capable
of
building
invariant
object
representations
from
impoverished
input
at

the
onset
of
vision.

Finally,
the
results
from
Chapters
5
&
6
demonstrate
that
newborn
chicks
can
build

representations
of
subordinate-‐level
objects.
In
Chapter
5,
we
tested
whether
this

automated
controlled-‐rearing
method
could
be
used
to
test
face
recognition
in
newborn

chicks.
Face
recognition
is
a
prototypical
example
of
subordinate-‐level
object
recognition

because
all
faces
share
the
same
basic
configuration.
We
found
that
newborn
chicks
can

distinguish
a
familiar
face
from
a
variety
of
unfamiliar
faces.
These
results
suggest
that

chicks
can
perform
fine-‐grained
differentiation,
discriminating
objects
that
share
the
same

general
configuration
of
features.
Chapter
6
extended
the
results
from
Chapter
5
by
testing

whether
newborn
chicks
can
perform
view-‐invariant
face
recognition.
We
found
that

newborn
chicks
can
build
a
view-‐invariant
representation
of
the
first
face
seen
in
life.

Impressively,
the
chicks
were
able
to
succeed
in
this
task
even
when
color
information
was

not
available
as
a
cue
for
recognition.
Taken
as
a
whole,
Chapters
2-‐6
demonstrate
that

newborn
chicks
are
capable
of
impressive
object
and
face
recognition
abilities
prior
to

extensive
visual
experience
with
objects.

127

Experiential
constraints
on
the
development
of
object
recognition

Although
invariant
object
recognition
can
emerge
rapidly
(within
the
first
days
of

life),
this
dissertation
also
provides
evidence
that
this
ability
does
not
develop

automatically.
Instead,
newborn
chicks
require
a
specific
type
of
visual
input
to
learn
to

recognize
objects
across
novel
viewing
situations.
While
Chapter
2
demonstrated
that

newborn
chicks
can
build
a
background-‐invariant
representation
of
the
first
object
seen
in

life,
Chapter
3
showed
that
this
ability
fails
to
develop
properly
when
newborn
chicks
are

deprived
of
visual
experience
with
objects
moving
on
patterned
backgrounds.
Specifically,

newborn
chicks
were
raised
with
a
single
object
rotating
either
on
no
background
(a
white

homogenous
screen)
or
on
multiple
patterned
backgrounds.
When
raised
with
an
object

moving
on
no
background,
the
chicks
generally
failed
to
recognize
the
object
across
novel

backgrounds
for
the
first
few
days
of
testing.
Conversely,
when
raised
with
an
object

moving
on
patterned
backgrounds,
the
chicks
successfully
recognized
the
object
across

novel
backgrounds
at
the
onset
of
testing.
Furthermore,
in
both
rearing
conditions,

recognition
performance
continued
to
increase
across
the
test
phase,
suggesting
that

background-‐invariant
object
recognition
continues
to
improve
as
newborns
acquire

greater
amounts
of
experience
with
objects
moving
on
backgrounds.

These
results
point
to
a
previously
unknown
constraint
on
the
development
of

object
segmentation
and
recognition:
newborns
need
experience
with
patterned

backgrounds
to
build
background-‐invariant
object
representations.
Prior
research
with

infants
and
patients
recovering
from
blindness
had
determined
that
object
motion
is
a

critical
cue
for
segmenting
objects;
however,
it
was
unknown
whether
object
motion
alone

is
sufficient
for
a
developing
visual
system
to
learn
how
to
segment
objects
from

128

backgrounds.
Our
results
demonstrate
that
object
motion
alone
is
not
sufficient
to
learn

how
to
segment
objects
from
the
background.
Experience
with
patterned
backgrounds
is

also
critical
for
learning
background-‐invariant
recognition.

Previous
studies
of
newborn
chicks
have
demonstrated
other
constraints
on
the

development
of
object
recognition.
For
instance,
we
have
found
that
objects
must
move

smoothly
(Wood
&
Wood,
under
review)
and
slowly
(Wood
&
Wood,
2016a)
during

encoding
for
newborns
to
build
view-‐invariant
representations
of
those
objects.
However,

all
of
these
studies
focused
on
basic-‐level
recognition.
No
studies
thus
far
have
investigated

whether
the
same
developmental
constraints
also
affect
other
levels
of
recognition.
To

address
this
question,
in
Chapter
6,
we
tested
whether
there
is
a
slowness
constraint
on
the

development
of
face
recognition.
We
used
faces
because
they
are
a
prototypical
example
of

subordinate-‐level
recognition.
Newborn
chicks
were
raised
for
the
first
week
of
life
with
a

single
face
that
rotated
back
and
forth
at
a
slow,
medium,
or
fast
speed.
The
same
pattern
of

results
emerged
for
faces
as
was
found
for
basic-‐level
objects.
Specifically,
newborn
chicks

that
were
raised
with
a
quickly
moving
face
built
face
representations
that
were
highly

sensitive
to
view-‐specific
information,
but
completely
insensitive
to
identity
information.

Conversely,
newborn
chicks
that
were
raised
with
a
slowly
moving
face
built
accurate
face

representations
that
were
highly
sensitive
to
identity
information,
but
completely

insensitive
to
viewpoint-‐specific
information.
Overall,
this
pattern
of
results
indicates
that

newborn
chicks
leverage
the
spatiotemporal
statistics
of
their
visual
environment
to
learn

about
the
external
world.

129

Future
Directions

These
results
raise
several
additional
questions
about
the
developmental
origins
of

object
recognition.
First,
what
are
the
object
features
extracted
by
newborn
visual
systems?

While
this
dissertation
presents
evidence
that
newborn
chicks
can
perform
background-‐

and
view-‐invariant
object
recognition,
these
results
do
not
necessarily
warrant
the
claim

that
newborns
can
build
geometric
three-‐dimensional
representations
of
whole
objects.

For
example,
newborn
chicks
could
rely
on
one
or
more
subfeatures
of
an
object
to
perform

recognition.
In
future
studies,
I
plan
to
study
newborn
visual
recognition
strategies
by

using
an
image
masking
technique
(Alemi-‐Neissi
et
al.,
2013;
Gosselin
&
Schyns,
2001).

This
technique
reveals
the
diagnostic
features
used
by
subjects
to
discriminate
objects.
To

date,
this
methodology
has
been
impossible
to
use
with
newborns
because
the
image

masking
technique
requires
presenting
hundreds
of
test
trials
and
adapting
the
stimuli
as
a

function
of
performance.
In
image
masking
studies,
an
opaque
mask
with
a
number
of

transparent
windows
(or
“bubbles”)
is
superimposed
on
a
visual
stimulus.
During
testing,

the
number
of
bubbles
is
adjusted
until
performance
reaches
a
predetermined
threshold.

Thus,
the
experimenter
must
test
each
subject
on
a
range
of
bubble
configurations,

determine
the
appropriate
number
of
bubbles
for
each
individual
subject,
and
then
modify

the
remaining
test
trials
to
utilize
the
correct
number
of
bubbles
for
that
subject.
While
the

current
automated
controlled-‐rearing
chambers
are
not
equipped
to
adjust
the
stimuli
as
a

function
of
the
chicks’
performance,
recent
methodological
developments
from
the
lab

should
enable
such
a
study
in
the
near
future.
Specifically,
the
lab
has
recently
developed

the
first
automated
virtual
reality
(VR)
system
for
testing
newborn
subjects;
these
VR

chambers
update
the
virtual
environment
in
response
to
subjects’
movements

130

(performance)
in
real
time.
With
this
new
advance,
it
is
now
possible
to
create
test
stimuli

that
interact
contingently
with
the
newborn
chicks.
Studies
testing
newborn
chicks
on

masked
images
would
provide
the
first
glimpse
into
the
features
used
by
newborn
visual

systems
to
recognize
objects.

Second,
how
does
the
newborn
brain
leverage
experience
with
backgrounds
to
learn

background-‐invariant
object
recognition?
In
Chapter
3,
we
found
that
newborn
chicks
were

impaired
at
background-‐invariant
object
recognition
when
they
did
not
have
experience

viewing
an
object
moving
across
a
patterned
background
prior
to
testing.
However,
the

exact
amount
and
type
of
background
experience
that
is
necessary
for
newborns
to
succeed

in
this
task
is
unknown.
Do
newborns
need
visual
experience
with
an
object
moving
across

a
patterned
background
or
is
viewing
a
moving
object
and
a
background
scene
separately

sufficient
for
background-‐invariant
recognition?
What
amount
or
type
of
background

features
is
sufficient
to
enable
background-‐invariant
recognition?
For
example,
do

background
images
need
to
vary
along
specific
ranges
of
spatial
frequency?
Addressing

these
questions
will
shed
light
on
the
mechanisms
of
object
segmentation
in
the
newborn

brain.

Third,
are
newborn
object
representations
invariant
to
other
transformations
in

appearance
(besides
backgrounds
and
viewpoints)?
For
example,
changes
in
lighting
can

create
massive
differences
in
low-‐level
features
such
as
hue
and
overall
brightness.
Mature

human
visual
systems
automatically
estimate
actual
reflectance
independent
of
the
lighting

conditions,
a
task
called
“lightness
constancy”
(Adelson,
2000).
Are
newborn
visual

systems
capable
of
lightness
constancy
at
the
onset
of
vision?
Are
slow
and
smooth
changes

in
lighting
important
for
developing
lightness
constancy
(akin
to
the
relationship
between

131

slow
and
smooth
motion
and
viewpoint
invariance)?
Do
changes
in
illumination
have

similar
effects
on
newborns’
perception
of
both
the
three-‐dimensional
structure
of
an

object
and
the
object’s
texture?

Finally,
what
neural
algorithms
give
rise
to
object
recognition
in
the
newborn
brain?

The
results
presented
here
provide
a
unique
glimpse
into
the
visuo-‐cognitive
abilities
of

the
newborn
brain,
but
the
computational
architecture
of
the
newborn
mind
remains

unknown.
Indeed,
modeling
the
visual
abilities
of
newborn
animals
is
a
unique
endeavor.

While
many
computational
models
of
object
recognition
exist,
the
goal
of
these
models
is

typically
either
(1)
to
optimize
overall
accuracy
of
performance
and/or
(2)
to
optimize
the

similarity
between
the
performance
of
the
models
and
adult
human
performance
(Borji
&

Itti,
2014).
However,
since
adults
have
had
years
of
experience
learning
about
the
visual

world,
models
designed
to
match
adult
vision
do
not
provide
information
about
the
initial

state
of
object
recognition,
nor
do
these
models
reveal
how
that
initial
state
is
shaped
by

visual
experience.

To
address
this
gap
in
the
literature,
I
have
begun
building
biologically-‐inspired

convolutional
neural
networks
that
learn
about
visual
input
using
unsupervised
temporal

learning
algorithms.
Since
an
object’s
identity
is
temporally
stable,
different
retinal
images

of
the
same
object
tend
to
be
contiguous
over
time.
Thus,
the
visual
system
might
build

invariant
representations
by
learning
the
spatiotemporal
statistics
produced
by

consecutive
retinal
images
of
an
object.
The
work
presented
in
my
dissertation
provides

preliminary
evidence
for
this
class
of
models.
My
ultimate
goal
is
to
identify
computational

models
that
produce
the
same
patterns
of
behavioral
performance
as
the
newborn
chicks

tested
in
our
controlled-‐rearing
experiments.

132

So
far,
this
has
proven
to
be
a
herculean
task.
Newborn
chicks’
visual
systems
are

generative
enough
to
build
object
representations
from
just
three
images
of
an
object

(Chapter
4),
but
flexible
enough
to
respond
to
the
spatiotemporal
information
provided
by

object
motion
(Chapter
6;
Wood,
2016;
Wood
et
al.,
2016;
Wood
&
Wood,
2016a;
Wood
&

Wood,
under
review).
I
have
found
that
models
using
unsupervised
temporal
learning

mechanisms
are
subject
to
the
same
constraints
on
object
recognition
as
newborn
chicks.

Namely,
unsupervised
temporal
learning
models
are
better
able
to
learn
about
objects
that

move
smoothly
and
slowly
over
time.
However,
while
these
models
are
promising
in
some

respects,
they
are
still
not
sufficient
to
account
for
the
impressive
generative
recognition

abilities
found
in
newborn
chicks.
Future
research
will
need
to
explore
whether
other

computational
mechanisms
(such
as
prior
connection
weights
and
recurrent
layers)
can

enable
models
to
learn
from
such
impoverished
visual
input.

In
summary,
these
studies
used
an
automated
controlled-‐rearing
method
to
study

the
development
of
object
and
face
recognition
in
newborns.
The
results
demonstrate
that

newborn
brains
contain
advanced
visual
processing
machinery:
newborn
visual
systems

can
build
abstract
representations
of
objects
and
faces
from
highly
impoverished
visual

input.

133

References

Adelson,
E.
H.
(2000).
Lightness
Perception
and
Lightness
Illusions.
In
M.
Gazzaniga
(Ed.),

The
New
Cognitive
Neurosciences
(2nd
ed.
ed.,
pp.
339-‐351).
Cambridge,
MA:
MIT

Press.

Alemi-‐Neissi,
A.,
Rosselli,
F.
B.,
&
Zoccolan,
D.
(2013).
Multifeatural
shape
processing
in
rats

engaged
in
invariant
visual
object
recognition.
Journal
of
Neuroscience,
33(14),

5939-‐5956.
doi:10.1523/JNEUROSCI.3629-‐12.2013

Arterberry,
M.
E.,
&
Yonas,
A.
(2000).
Perception
of
three-‐dimensional
shape
specified
by

optic
flow
by
8-‐week-‐old
infants.
Perception
and
Psychophysics,
62(3),
550-‐556.

doi:Doi
10.3758/Bf03212106

Bateson,
P.
(2000).
What
must
be
known
in
order
to
understand
imprinting?
In
C.
Heyes
&

L.
Huber
(Eds.),
The
evolution
of
cognition
(pp.
85-‐102).
Cambridge,
MA:
The
MIT

Press.

Bellen,
H.
J.,
Tong,
C.,
&
Tsuda,
H.
(2010).
100
years
of
Drosophila
research
and
its
impact

on
vertebrate
neuroscience:
a
history
lesson
for
the
future.
Nature
Reviews

Neuroscience,
11(7),
514-‐522.
doi:10.1038/nrn2839

Berkeley,
G.
(1910).
A
new
theory
of
vision,
and
other
writings.
London:
Dent.

Biederman,
I.
(1987).
Recognition-‐by-‐Components
-‐
a
Theory
of
Human
Image

Understanding.
Psychological
Review,
94(2),
115-‐147.
doi:Doi
10.1037/0033-‐
295x.94.2.115

Bogale,
B.
A.,
Aoyama,
M.,
&
Sugita,
S.
(2011).
Categorical
learning
between
'male'
and

'female'
photographic
human
faces
in
jungle
crows
(Corvus
macrorhynchos).

Behavioural
Processes,
86(1),
109-‐118.
doi:Doi
10.1016/J.Beproc.2010.10.002

Bolhuis,
J.
J.
(1999).
Early
learning
and
the
development
of
filial
preferences
in
the
chick.

Behavioural
Brain
Research,
98(2),
245-‐252.

Borji,
A.,
&
Itti,
L.
(2014).
Human
vs.
computer
in
scene
and
object
recognition.
Paper

presented
at
the
Proceedings
of
the
IEEE
Conference
on
Computer
Vision
and

Pattern
Recognition.

Brenner,
S.
(1974).
The
genetics
of
Caenorhabditis
elegans.
Genetics,
77(1),
71-‐94.

Bruce,
V.,
Campbell,
R.
N.,
Doherty-‐Sneddon,
G.,
Import,
A.,
Langton,
S.,
McAuley,
S.,
&

Wright,
R.
(2000).
Testing
face
processing
skills
in
children.
British
Journal
of

Developmental
Psychology,
18,
319-‐333.

Bulf,
H.,
Johnson,
S.
P.,
&
Valenza,
E.
(2011).
Visual
statistical
learning
in
the
newborn

infant.
Cognition,
121(1),
127-‐132.
doi:10.1016/j.cognition.2011.06.010

134

Butler,
A.
B.
(1994).
The
evolution
of
the
dorsal
pallium
in
the
telencephalon
of
amniotes:

cladistic
analysis
and
a
new
hypothesis.
Brain
Research
Reviews,
19(1),
66-‐101.

Carey,
S.
(1992).
Becoming
a
Face
Expert.
Philosophical
Transactions
of
the
Royal
Society
of

London
Series
B-‐Biological
Sciences,
335(1273),
95-‐103.

doi:10.1098/Rstb.1992.0012

Carey,
S.
(2009).
The
origin
of
concepts.
Oxford
;
New
York:
Oxford
University
Press.

Carey,
S.,
&
Diamond,
R.
(1977).
From
Piecemeal
to
Configurational
Representation
of

Faces.
Science,
195(4275),
312-‐314.
doi:Doi
10.1126/Science.831281

Chiandetti,
C.,
&
Vallortigara,
G.
(2011).
Intuitive
physical
reasoning
about
occluded
objects

by
inexperienced
chicks.
Proceedings
of
the
Royal
Society
B-‐Biological
Sciences,

278(1718),
2621-‐2627.
doi:10.1098/rspb.2010.2381

Collignon,
O.,
Dormal,
G.,
de
Heering,
A.,
Lepore,
F.,
Lewis,
T.
L.,
&
Maurer,
D.
(2015).
Long-‐
Lasting
Crossmodal
Cortical
Reorganization
Triggered
by
Brief
Postnatal
Visual

Deprivation.
Current
Biology,
25(18),
2379-‐2383.

Cox,
D.
D.,
Meier,
P.,
Oertelt,
N.,
&
DiCarlo,
J.
J.
(2005).
'Breaking'
position-‐invariant
object

recognition.
Nature
Neuroscience,
8(9),
1145-‐1147.
doi:10.1038/nn1519

Damasio,
A.
R.,
Damasio,
H.,
&
Vanhoesen,
G.
W.
(1982).
Prosopagnosia
-‐
Anatomic
Basis

and
Behavioral
Mechanisms.
Neurology,
32(4),
331-‐341.

de
Haan,
M.,
Johnson,
M.
H.,
Maurer,
D.,
&
Perrett,
D.
I.
(2001).
Recognition
of
individual

faces
and
average
face
prototypes
by
1-‐and
3-‐month-‐old
infants.
Cognitive

Development,
16(2),
659-‐678.

Dell,
A.
I.,
Bender,
J.
A.,
Branson,
K.,
Couzin,
I.
D.,
de
Polavieja,
G.
G.,
Noldus,
L.
P.,
.
.
.
Brose,
U.

(2014).
Automated
image-‐based
tracking
and
its
application
in
ecology.
Trends
in

Ecology
&
Evolution.
doi:10.1016/j.tree.2014.05.004

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

Journal
of
Experimental
Psychology-‐General,
115(2),
107-‐117.

DiCarlo,
J.
J.,
Zoccolan,
D.,
&
Rust,
N.
C.
(2012).
How
does
the
brain
solve
visual
object

recognition?
Neuron,
73(3),
415-‐434.
doi:10.1016/J.Neuron.2012.01.010

Dugas-‐Ford,
J.,
Rowell,
J.
J.,
&
Ragsdale,
C.
W.
(2012).
Cell-‐type
homologies
and
the
origins
of

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

America,
109(42),
16974-‐16979.
doi:10.1073/Pnas.1204773109

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

statically
bound
visual
representations.
Cognitive
Science,
27(1),
73-‐109.

135

Espinosa,
J.
S.,
&
Stryker,
M.
P.
(2012).
Development
and
plasticity
of
the
primary
visual

cortex.
Neuron,
75(2),
230-‐249.

Farah,
M.
J.,
Wilson,
K.
D.,
Drain,
M.,
&
Tanaka,
J.
N.
(1998).
What
is
"special"
about
face

perception?
Psychological
Review,
105(3),
482-‐498.

Frank,
M.
C.,
Vul,
E.,
&
Johnson,
S.
P.
(2009).
Development
of
infants'
attention
to
faces

during
the
first
year.
Cognition,
110(2),
160-‐170.

doi:10.1016/j.cognition.2008.11.010

Gavornik,
J.
P.,
&
Bear,
M.
F.
(2014).
Learned
spatiotemporal
sequence
recognition
and

prediction
in
primary
visual
cortex.
Nature
Neuroscience,
17(5),
732.

doi:10.1038/Nn.3683

Gibson,
B.
M.,
Wasserman,
E.
A.,
Gosselin,
F.,
&
Schyns,
P.
G.
(2005).
Applying
bubbles
to

localize
features
that
control
pigeons'
visual
discrimination
behavior.
Journal
of

Experimental
Psychology:
Animal
Behavior
Processes,
31(3),
376-‐382.

doi:10.1037/0097-‐7403.31.3.376

Goldman,
J.
G.,
&
Wood,
J.
N.
(2015).
An
automated
controlled-‐rearing
method
for
studying

the
origins
of
movement
recognition
in
newly
hatched
chicks.
Animal
Cognition,

18(3),
723-‐731.
doi:10.1007/s10071-‐015-‐0839-‐3

Gosselin,
F.,
&
Schyns,
P.
G.
(2001).
Bubbles:
a
technique
to
reveal
the
use
of
information
in

recognition
tasks.
Vision
Research,
41(17),
2261-‐2271.
doi:10.1016/S0042-‐
6989(01)00097-‐9

Hasselmo,
M.
E.,
Rolls,
E.
T.,
Baylis,
G.
C.,
&
Nalwa,
V.
(1989).
Object-‐Centered
Encoding
by

Face-‐Selective
Neurons
in
the
Cortex
in
the
Superior
Temporal
Sulcus
of
the

Monkey.
Experimental
Brain
Research,
75(2),
417-‐429.

Helmholtz,
H.
v.,
&
Southall,
J.
P.
C.
(1924).
Helmholtz's
treatise
on
physiological
optics.

Rochester,
N.Y.:
The
Optical
Society
of
America.

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

Shape
and
Color.
Proceedings
of
the
Royal
Society
of
London
Series
B-‐Biological

Sciences,
261(1362),
367-‐373.

Hill,
H.,
Schyns,
P.
G.,
&
Akamatsu,
S.
(1997).
Information
and
viewpoint
dependence
in
face

recognition.
Cognition,
62(2),
201-‐222.
doi:10.1016/S0010-‐0277(96)00785-‐8

Hochberg,
J.
E.
(1978).
Perception
(2d
ed.).
Englewood
Cliffs,
N.J.:
Prentice-‐Hall.

Horn,
G.
(2004).
Pathways
of
the
past:
The
imprint
of
memory.
Nature
Reviews:

Neuroscience,
5(2),
108-‐U113.

136

Ishikawa,
A.
W.,
Komatsu,
Y.,
&
Yoshimura,
Y.
(2014).
Experience-‐Dependent
Emergence
of

Fine-‐Scale
Networks
in
Visual
Cortex.
Journal
of
Neuroscience,
34(37),
12576-‐12586.

doi:10.1523/Jneurosci.1346-‐14.2014

Jarvis,
E.
D.,
Gunturkun,
O.,
Bruce,
L.,
Csillag,
A.,
Karten,
H.,
Kuenzel,
W.,
.
.
.
Butler,
A.
B.

(2005).
Avian
brains
and
a
new
understanding
of
vertebrate
brain
evolution.
Nature

Reviews:
Neuroscience,
6(2),
151-‐159.
doi:10.1038/nrn1606

Johnson,
M.
H.,
Dziurawiec,
S.,
Ellis,
H.,
&
Morton,
J.
(1991).
Newborns
preferential
tracking

of
face-‐like
stimuli
and
its
subsequent
decline.
Cognition,
40(1-‐2),
1-‐19.

Johnson,
S.
P.,
Amso,
D.,
&
Slemmer,
J.
A.
(2003).
Development
of
object
concepts
in
infancy:

Evidence
for
early
learning
in
an
eye-‐tracking
paradigm.
Proceedings
of
the
National

Academy
of
Sciences,
100(18),
10568-‐10573.

Johnson,
S.
P.,
&
Aslin,
R.
N.
(1995).
Perception
of
Object
Unity
in
2-‐Month-‐Old
Infants.

Developmental
Psychology,
31(5),
739-‐745.
doi:10.1037/0012-‐1649.31.5.739

Johnson,
S.
P.,
Bremner,
J.
G.,
Slater,
A.
M.,
Mason,
U.
C.,
&
Foster,
K.
(2002).
Young
infants'

perception
of
unity
and
form
in
occlusion
displays.
Journal
of
Experimental
Child

Psychology,
81(3),
358-‐374.
doi:10.1006/jecp.2002.2657

Johnson,
S.
P.,
Schwarzer,
G.,
&
Leder,
H.
(2003).
Development
of
fragmented
versus
holistic

object
perception.
The
development
of
face
processing,
3-‐17.

Kandel,
E.
R.
(2007).
In
search
of
memory:
The
emergence
of
a
new
science
of
mind:
WW

Norton
&
Company.

Kant,
I.,
Guyer,
P.,
&
Wood,
A.
W.
(1998).
Critique
of
pure
reason.
Cambridge
;
New
York:

Cambridge
University
Press.

Karten,
H.
J.
(1969).
The
organization
of
the
avian
telencephalon
and
some
speculations
on

the
phylogeny
of
the
amniote
telencephalon.
Annals
of
the
New
York
Academy
of

Sciences,
167(1),
164-‐179.

Karten,
H.
J.
(1991).
Homology
and
evolutionary
origins
of
the'neocortex'.
Brain,
Behavior

and
Evolution,
38(4-‐5),
264-‐272.

Karten,
H.
J.
(1997).
Evolutionary
developmental
biology
meets
the
brain:
the
origins
of

mammalian
cortex.
Proceedings
of
the
National
Academy
of
Sciences,
94(7),
2800-‐
2804.

Karten,
H.
J.
(2013).
Neocortical
evolution:
neuronal
circuits
arise
independently
of

lamination.
Current
Biology,
23(1),
R12-‐15.
doi:10.1016/j.cub.2012.11.013

Karten,
H.
J.,
&
Shimizu,
T.
(1989).
The
origins
of
neocortex:
connections
and
lamination
as

distinct
events
in
evolution.
Journal
of
Cognitive
Neuroscience,
1(4),
291-‐301.

137

Kellman,
P.
J.
(1984).
Perception
of
three-‐dimensional
form
by
human
infants.
Perception

and
Psychophysics,
36(4),
353-‐358.

Kellman,
P.
J.,
&
Short,
K.
R.
(1987).
Development
of
three-‐dimensional
form
perception.

Journal
of
Experimental
Psychology:
Human
Perception
and
Performance,
13(4),
545.

Kellman,
P.
J.,
&
Spelke,
E.
S.
(1983).
Perception
of
Partly
Occluded
Objects
in
Infancy.

Cognitive
Psychology,
15(4),
483-‐524.
doi:10.1016/0010-‐0285(83)90017-‐8

Kellman,
P.
J.,
Spelke,
E.
S.,
&
Short,
K.
R.
(1986).
Infant
Perception
of
Object
Unity
from

Translatory
Motion
in
Depth
and
Vertical
Translation.
Child
Development,
57(1),
72-‐
86.
doi:10.1111/j.1467-‐8624.1986.tb00008.x

Kelly,
D.
J.,
Quinn,
P.
C.,
Slater,
A.
M.,
Lee,
K.,
Ge,
L.
Z.,
&
Pascalis,
O.
(2007).
The
other-‐race

effect
develops
during
infancy
-‐
Evidence
of
perceptual
narrowing.
Psychological

Science,
18(12),
1084-‐1089.
doi:10.1111/j.1467-‐9280.2007.02029.x

Kestenbaum,
R.,
Termine,
N.,
&
Spelke,
E.
S.
(1987).
Perception
of
Objects
and
Object

Boundaries
by
3-‐Month-‐Old
Infants.
British
Journal
of
Developmental
Psychology,
5,

367-‐383.

Kirkham,
N.
Z.,
Slemmer,
J.
A.,
&
Johnson,
S.
P.
(2002).
Visual
statistical
learning
in
infancy:

evidence
for
a
domain
general
learning
mechanism.
Cognition,
83(2),
B35-‐42.

Ko,
H.,
Mrsic-‐Flogel,
T.
D.,
&
Hofer,
S.
B.
(2014).
Emergence
of
feature-‐specific
connectivity

in
cortical
microcircuits
in
the
absence
of
visual
experience.
Journal
of
Neuroscience,

34(29),
9812-‐9816.

Koffka,
K.
(1935).
Principles
of
Gestalt
psychology.
London:
K.
Paul,
Trench.

Köhler,
W.
(1929).
Gestalt
psychology.
New
York,:
H.
Liveright.

Krizhevsky,
A.,
Sutskever,
I.,
&
Hinton,
G.
E.
(2012).
Imagenet
classification
with
deep

convolutional
neural
networks.
Paper
presented
at
the
Advances
in
Neural

Information
Processing
Systems.

Leibo,
J.
Z.,
Mutch,
J.,
&
Poggio,
T.
(2011).
Why
The
Brain
Separates
Face
Recognition
From

Object
Recognition.
NIPS,
711-‐719.

Li,
N.,
&
DiCarlo,
J.
J.
(2008).
Unsupervised
natural
experience
rapidly
alters
invariant
object

representation
in
visual
cortex.
Science,
321(5895),
1502-‐1507.

Li,
N.,
&
DiCarlo,
J.
J.
(2010).
Unsupervised
natural
visual
experience
rapidly
reshapes
size-‐
invariant
object
representation
in
inferior
temporal
cortex.
Neuron,
67(6),
1062-‐
1075.
doi:10.1016/j.neuron.2010.08.029

Logothetis,
N.
K.,
&
Sheinberg,
D.
L.
(1996).
Visual
object
recognition.
Annual
Review
of

Neuroscience,
19,
577-‐621.
doi:10.1146/annurev.ne.19.030196.003045

138

Loken,
E.,
&
Gelman,
A.
(2017).
Measurement
error
and
the
replication
crisis.
Science,

355(6325),
584-‐585.

Maidenbaum,
S.,
Abboud,
S.,
&
Amedi,
A.
(2014).
Sensory
substitution:
Closing
the
gap

between
basic
research
and
widespread
practical
visual
rehabilitation.
Neuroscience

and
Biobehavioral
Reviews,
41,
3-‐15.

Martinho,
A.,
&
Kacelnik,
A.
(2016).
Ducklings
imprint
on
the
relational
concept
of
"same
or

different".
Science,
353(6296),
286-‐288.
doi:10.1126/science.aaf4247

Mascalzoni,
E.,
Regolin,
L.,
&
Vallortigara,
G.
(2010).
Innate
sensitivity
for
self-‐propelled

causal
agency
in
newly
hatched
chicks.
Proceedings
of
the
National
Academy
of

Sciences
of
the
United
States
of
America,
107(9),
4483-‐4485.

doi:10.1073/pnas.0908792107

Masquelier,
T.,
&
Thorpe,
S.
J.
(2007).
Unsupervised
learning
of
visual
features
through

spike
timing
dependent
plasticity.
PLoS
Computational
Biology,
3(2),
e31.

doi:10.1371/journal.pcbi.0030031

McKyton,
A.,
Ben-‐Zion,
I.,
Doron,
R.,
&
Zohary,
E.
(2015).
The
Limits
of
Shape
Recognition

following
Late
Emergence
from
Blindness.
Current
Biology,
25(18),
2373-‐2378.

doi:10.1016/j.cub.2015.06.040

Medina,
L.,
&
Reiner,
A.
(2000).
Do
birds
possess
homologues
of
mammalian
primary
visual,

somatosensory
and
motor
cortices?
Trends
in
Neurosciences,
23(1),
1-‐12.

Mondloch,
C.
J.,
Le
Grand,
R.,
&
Maurer,
D.
(2010).
Development
of
expertise
in
face

recognition.
In
I.
Gauthier,
M.
J.
Tarr,
&
D.
Bub
(Eds.),
Perceptual
expertise;
Bridging

brain
and
behavior
(pp.
67-‐106):
Oxford
University
Press.

Morales,
B.,
Choi,
S.-‐Y.,
&
Kirkwood,
A.
(2002).
Dark
rearing
alters
the
development
of

GABAergic
transmission
in
visual
cortex.
Journal
of
Neuroscience,
22(18),
8084-‐
8090.

Moses,
Y.,
Ullman,
S.,
&
Edelman,
S.
(1996).
Generalization
to
novel
images
in
upright
and

inverted
faces.
Perception,
25(4),
443-‐461.
doi:10.1068/p250443

Needham,
A.
(2000).
Improvements
in
Object
Exploration
Skills
May
Facilitate
the

Development
of
Object
Segregation
in
Early
Infancy.
Journal
of
Cognition
and

Development,
1(2),
131-‐156.
doi:10.1207/S15327647jcd010201

Needham,
A.,
&
Ormsbee,
S.
M.
(2003).
The
development
of
object
segregation
during
the

first
year
of
life.
In
R.
Kimchi,
M.
Behrmann,
&
C.
R.
Olson
(Eds.),
Perceptual

organization
in
vision:
Behavioral
and
neural
perspectives
(pp.
205-‐232).
Mahwah,

NJ:
Lawrence
Erlbaum
Associates
Publishers.

139

Olshausen,
B.
A.,
&
Field,
D.
J.
(1996).
Emergence
of
simple-‐cell
receptive
field
properties
by

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

Ostrovsky,
Y.,
Meyers,
E.,
Ganesh,
S.,
Mathur,
U.,
&
Sinha,
P.
(2009).
Visual
Parsing
After

Recovery
From
Blindness.
Psychological
Science,
20(12),
1484-‐1491.

Owsley,
C.
(1983).
The
role
of
motion
in
infants'
perception
of
solid
shape.
Perception,

12(6),
707-‐717.

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

Recognition
by
Neonates
-‐
a
Replication
and
an
Extension.
Infant
Behavior
&

Development,
18(1),
79-‐85.

Pettigrew,
J.
D.,
&
Konishi,
M.
(1976).
Neurons
selective
for
orientation
and
binocular

disparity
in
the
visual
Wulst
of
the
barn
owl
(Tyto
alba).
Science,
193(4254),
675-‐
678.

Piaget,
J.
(1952).
The
origins
of
intelligence
in
children.
New
York:
International
Universities

Press.

Pinto,
N.,
Cox,
D.
D.,
&
DiCarlo,
J.
J.
(2008).
Why
is
real-‐world
visual
object
recognition
hard?

PLoS
Computational
Biology,
4(1),
27.
doi:10.1371/journal.pcbi.0040027.sg004

Regolin,
L.,
Rugani,
R.,
Stancher,
G.,
&
Vallortigara,
G.
(2011).
Spontaneous
discrimination
of

possible
and
impossible
objects
by
newly
hatched
chicks.
Biology
Letters,
7(5),
654-‐
657.
doi:10.1098/rsbl.2011.0051

Regolin,
L.,
&
Vallortigara,
G.
(1995).
Perception
of
Partly
Occluded
Objects
by
Young

Chicks.
Perception
and
Psychophysics,
57(7),
971-‐976.
doi:10.3758/Bf03205456

Reid,
T.
(1764).
An
Inquiry
into
the
Human
Mind:
in
the
version
by
Jonathan
Bennett

presented
at
http://www.earlymoderntexts.com.

Reiner,
A.,
Yamamoto,
K.,
&
Karten,
H.
J.
(2005).
Organization
and
evolution
of
the
avian

forebrain.
The
Anatomical
Record,
287(1),
1080-‐1102.

Roder,
B.,
Ley,
P.,
Shenoy,
B.
H.,
Kekunnaya,
R.,
&
Bottari,
D.
(2013).
Sensitive
periods
for

the
functional
specialization
of
the
neural
system
for
human
face
processing.

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

110(42),
16760-‐16765.
doi:10.1073/pnas.1309963110

Rolls,
E.
T.
(2000).
Functions
of
the
primate
temporal
lobe
cortical
visual
areas
in
invariant

visual
object
and
face
recognition.
Neuron,
27(2),
205-‐218.
doi:10.1016/s0896-‐
6273(00)00030-‐1

140

Rosa-‐Salva,
O.,
Farroni,
T.,
Regolin,
L.,
Vallortigara,
G.,
&
Johnson,
M.
H.
(2011).
The

Evolution
of
Social
Orienting:
Evidence
from
Chicks
(Gallus
gallus)
and
Human

Newborns.
PloS
One,
6(4).

Rosa-‐Salva,
O.,
Grassi,
M.,
Lorenzi,
E.,
Regolin,
L.,
&
Vallortigara,
G.
(2016).
Spontaneous

preference
for
visual
cues
of
animacy
in
naive
domestic
chicks:
The
case
of
speed

changes.
Cognition,
157,
49-‐60.
doi:10.1016/j.cognition.2016.08.014

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
Science,
13(4),
565-‐577.

doi:10.1111/J.1467-‐7687.2009.00914.X

Rosa-‐Salva,
O.,
Regolin,
L.,
&
Vallortigara,
G.
(2012).
Inversion
of
contrast
polarity
abolishes

spontaneous
preferences
for
face-‐like
stimuli
in
newborn
chicks.
Behavioural
Brain

Research,
228(1),
133-‐143.

Ruff,
H.
A.
(1978).
Infant
recognition
of
the
invariant
form
of
objects.
Child
Development,

293-‐306.

Saffran,
J.
R.,
Aslin,
R.
N.,
&
Newport,
E.
L.
(1996).
Statistical
learning
by
8-‐month-‐old

infants.
Science,
274(5294),
1926-‐1928.

Said,
C.
P.,
&
Todorov,
A.
(2011).
A
Statistical
Model
of
Facial
Attractiveness.
Psychological

Science,
22(9),
1183-‐1190.
doi:10.1177/0956797611419169

Saini,
K.,
&
Leppelsack,
H.
J.
(1981).
Cell
types
of
the
auditory
caudomedial
neostriatum
of

the
starling
(Sturnus
vulgaris).
Journal
of
Comparative
Neurology,
198(2),
209-‐229.

Shanahan,
M.,
Bingman,
V.
P.,
Shimizu,
T.,
Wild,
M.,
&
Gunturkun,
O.
(2013).
Large-‐scale

network
organization
in
the
avian
forebrain:
a
connectivity
matrix
and
theoretical

analysis.
Frontiers
in
Computational
Neuroscience,
7,
89.

doi:10.3389/fncom.2013.00089

Simmons,
J.
P.,
Nelson,
L.
D.,
&
Simonsohn,
U.
(2011).
False-‐positive
psychology:

Undisclosed
flexibility
in
data
collection
and
analysis
allows
presenting
anything
as

significant.
Psychological
Science,
22(11),
1359-‐1366.

Sinha,
P.
(2013).
Once
blind
and
now
they
see.
Scientific
American,
309(1),
48-‐55.

Soska,
K.
C.,
&
Johnson,
S.
P.
(2008).
Development
of
three-‐dimensional
object
completion

in
infancy.
Child
Development,
79(5),
1230-‐1236.
doi:10.1111/j.1467-‐
8624.2008.01185.x

Soto,
F.
A.,
Siow,
J.
Y.,
&
Wasserman,
E.
A.
(2012).
View-‐invariance
learning
in
object

recognition
by
pigeons
depends
on
error-‐driven
associative
learning
processes.

Vision
Research,
62,
148-‐161.

141

Spelke,
E.
S.
(1990).
Principles
of
object
perception.
Cognitive
Science,
14(1),
29-‐56.

Spelke,
E.
S.,
&
Kinzler,
K.
D.
(2007).
Core
knowledge.
Developmental
Science,
10(1),
89-‐96.

Srihasam,
K.,
Mandeville,
J.
B.,
Morocz,
I.
A.,
Sullivan,
K.
J.,
&
Livingstone,
M.
S.
(2012).

Behavioral
and
Anatomical
Consequences
of
Early
versus
Late
Symbol
Training
in

Macaques.
Neuron,
73(3),
608-‐619.

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

Proceedings
of
the
National
Academy
of
Sciences,
105(1),
394-‐398.

Tafazoli,
S.,
Di
Filippo,
A.,
&
Zoccolan,
D.
(2012).
Transformation-‐tolerant
object
recognition

in
rats
revealed
by
visual
priming.
Journal
of
Neuroscience,
32(1),
21-‐34.

doi:10.1523/JNEUROSCI.3932-‐11.2012

Tanaka,
K.
(1996).
Inferotemporal
cortex
and
object
vision.
Annual
Review
of
Neuroscience,

19,
109-‐139.
doi:10.1146/Annurev.Ne.19.030196.000545

Tarr,
M.
J.,
&
Gauthier,
I.
(2000).
FFA:
a
flexible
fusiform
area
for
subordinate-‐level
visual

processing
automatized
by
expertise.
Nature
Neuroscience,
3(8),
764-‐769.

doi:10.1038/77666

Troje,
N.
F.,
Huber,
L.,
Loidolt,
M.,
Aust,
U.,
&
Fieder,
M.
(1999).
Categorical
learning
in

pigeons:
the
role
of
texture
and
shape
in
complex
static
stimuli.
Vision
Research,

39(2),
353-‐366.
doi:10.1016/S0042-‐6989(98)00153-‐9

Vallortigara,
G.
(2012).
Core
knowledge
of
object,
number,
and
geometry:
A
comparative

and
neural
approach.
Cognitive
Neuropsychology,
29(1-‐2),
213-‐236.

Vallortigara,
G.,
Regolin,
L.,
&
Marconato,
F.
(2005).
Visually
inexperienced
chicks
exhibit

spontaneous
preference
for
biological
motion
patterns.
PLoS
Biology,
3(7),
1312-‐
1316.
doi:10.1371/journal.pbio.0030208

Vallortigara,
G.,
Regolin,
L.,
Rigoni,
M.,
&
Zanforlin,
M.
(1998).
Delayed
search
for
a

concealed
imprinted
object
in
the
domestic
chick.
Animal
Cognition,
1(1),
17-‐24.

Wallis,
G.
(2013).
Toward
a
unified
model
of
face
and
object
recognition
in
the
human

visual
system.
Frontiers
in
Psychology,
4.
doi:10.3389/Fpsyg.2013.00497

Wallis,
G.,
&
Bülthoff,
H.
H.
(2001).
Effects
of
temporal
association
on
recognition
memory.

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

98(8),
4800-‐4804.
doi:10.1073/pnas.071028598

Wallis,
G.,
&
Rolls,
E.
T.
(1997).
Invariant
face
and
object
recognition
in
the
visual
system.

Progress
in
Neurobiology,
51(2),
167-‐194.

142

Wang,
Y.,
Brzozowska-‐Prechtl,
A.,
&
Karten,
H.
J.
(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.
doi:10.1073/pnas.1006645107

Wasserman,
E.,
&
Biederman,
I.
(2012).
Recognition-‐by-‐components:
A
bird's
eye
view.
In

O.
F.
Lazareva,
T.
Shimizu,
&
E.
A.
Wasserman
(Eds.),
How
Animals
See
the
World:

Comparative
Behavior,
Biology
and
Evolution
of
Vision
(pp.
191-‐216):
Oxford

University
Press.

Weigelt,
S.,
Koldewyn,
K.,
Dilks,
D.
D.,
Balas,
B.,
McKone,
E.,
&
Kanwisher,
N.
(2014).
Domain-‐
specific
development
of
face
memory
but
not
face
perception.
Developmental

Science,
17(1),
47-‐58.
doi:10.1111/desc.12089

Wilkinson,
N.,
Paikan,
A.,
Gredeback,
G.,
Rea,
F.,
&
Metta,
G.
(2014).
Staring
us
in
the
face?

An
embodied
theory
of
innate
face
preference.
Developmental
Science,
17(6),
809-‐
825.
doi:10.1111/desc.12159

Wiskott,
L.,
&
Sejnowski,
T.
J.
(2002).
Slow
feature
analysis:
Unsupervised
learning
of

invariances.
Neural
Computation,
14(4),
715-‐770.

doi:10.1162/089976602317318938

Wood,
J.
N.
(2013).
Newborn
chickens
generate
invariant
object
representations
at
the

onset
of
visual
object
experience.
Proceedings
of
the
National
Academy
of
Sciences
of

the
United
States
of
America,
110(34),
14000-‐14005.

Wood,
J.
N.
(2014).
Newly
hatched
chicks
solve
the
visual
binding
problem.
Psychological

Science,
25(7),
1475-‐1481.
doi:10.1177/0956797614528955

Wood,
J.
N.
(2015).
Characterizing
the
information
content
of
a
newly
hatched
chick's
first

visual
object
representation.
Developmental
Science,
18(2),
194-‐205.

doi:10.1111/desc.12198

Wood,
J.
N.
(2016).
A
smoothness
constraint
on
the
development
of
object
recognition.

Cognition,
153,
140-‐145.

Wood,
J.
N.,
Prasad,
A.,
Goldman,
J.
G.,
&
Wood,
S.
M.
W.
(2016).
Enhanced
learning
of

natural
visual
sequences
in
newborn
chicks.
Animal
Cognition,
19(4),
835–845.

Wood,
J.
N.,
&
Wood,
S.
M.
W.
(2016a).
The
development
of
newborn
object
recognition
in

fast
and
slow
visual
worlds.
Proceedings
of
the
Royal
Society
B-‐Biological
Sciences,

283(1829),
20160166.

Wood,
J.
N.,
&
Wood,
S.
M.
W.
(2016b).
Measuring
the
speed
of
newborn
object
recognition

in
controlled
visual
worlds.
Developmental
Science.
doi:10.1111/desc.12470

143

Wood,
S.
M.
W.,
&
Wood,
J.
N.
(2015a).
A
chicken
model
for
studying
the
emergence
of

invariant
object
recognition.
Frontiers
in
Neural
Circuits,
9,
7.

doi:10.3389/fncir.2015.00007

Wood,
S.
M.
W.,
&
Wood,
J.
N.
(2015b).
Face
recognition
in
newly
hatched
chicks
at
the

onset
of
vision.
Journal
of
Experimental
Psychology:
Animal
Learning
and
Cognition,

41(2),
206.

Wyss,
R.,
Konig,
P.,
&
Verschure,
P.
F.
M.
J.
(2006).
A
model
of
the
ventral
visual
system

based
on
temporal
stability
and
local
memory.
PLoS
Biology,
4(5),
836-‐843.

doi:ARTN
e12010.1371/journal.pbio.0040120

Xu,
F.
(2007).
Sortal
concepts,
object
individuation,
and
language.
Trends
in
Cognitive

Sciences,
11(9),
400-‐406.

Yamins,
D.
L.
K.,
Hong,
H.,
Cadieu,
C.
F.,
Solomon,
E.
A.,
Seibert,
D.,
&
DiCarlo,
J.
J.
(2014).

Performance-‐optimized
hierarchical
models
predict
neural
responses
in
higher

visual
cortex.
Proceedings
of
the
National
Academy
of
Sciences
of
the
United
States
of

America,
111(23),
8619-‐8624.
doi:10.1073/pnas.1403112111

Yang,
J.
L.,
Kanazawa,
S.,
Yamaguchi,
M.
K.,
&
Motoyoshi,
I.
(2015).
Pre-‐constancy
Vision
in

Infants.
Current
Biology,
25(24),
3209-‐3212.
doi:10.1016/j.cub.2015.10.053

Yue,
X.
M.,
Biederman,
I.,
Mangini,
M.
C.,
von
der
Malsburg,
C.,
&
Amir,
O.
(2012).
Predicting

the
psychophysical
similarity
of
faces
and
non-‐face
complex
shapes
by
image-‐based

measures.
Vision
Research,
55,
41-‐46.
doi:10.1016/j.visres.2011.12.012

Zoccolan,
D.
(2015).
Invariant
visual
object
recognition
and
shape
processing
in
rats.

Behavioural
Brain
Research,
285,
10-‐33.
doi:10.1016/j.bbr.2014.12.053

Zoccolan,
D.,
Oertelt,
N.,
DiCarlo,
J.
J.,
&
Cox,
D.
D.
(2009).
A
rodent
model
for
the
study
of

invariant
visual
object
recognition.
Proceedings
of
the
National
Academy
of
Sciences

of
the
United
States
of
America,
106(21),
8748-‐8753.
doi:10.1073/pnas.0811583106

Abstract (if available)

Linked assets