Transcript www.tech.plym.ac.uk

Are colour categories innate or learned?
Insights from computational modelling
Tony Belpaeme
Artificial Intelligence Lab
Vrije Universiteit Brussel
Tony Belpaeme
VUB AI-lab
Situating the research
• Artificial Life modelling
– Uses computer simulation
– Investigates particular natural phenomena
– Provides theories which are to be referred back to
other disciplines
– Allows investigation of phenomena where
observational disciplines fall short.
Tony Belpaeme
VUB AI-lab
Perceptual categories
• The origins of perceptual categories
– Facial expressions
– Odour
– Colour
• Debate on the origins of perceptual categories
Tony Belpaeme
VUB AI-lab
Three positions
1. Genetic determinism (or nativism)
–
–
–
Perceptual categories, among others, are innate.
Either directly, or indirectly through other innate
mechanisms.
Chomsky, Jackendoff, Fodor, Pinker.
2. Empiricism
–
–
–
Perceptual categories are learned.
Through interaction between the individual and its
environment.
Elman, Piaget.
3. Culturalism
–
–
–
Tony Belpaeme
VUB AI-lab
Perceptual categories are learned.
Through social (linguistic) interaction with other
individuals and a shared environment.
Whorf, Tomasello, Davidoff.
Colour categories
• Case study for this work:
the origins of colour categories
• Why colour categories?
– Well-documented field
Anthropology, psychology, cognitive science, neurophysiology,
physics, philosophy, …
–
–
–
–
Well-known field
Tightly defined domain
Controversial
Easy to relate to
Tony Belpaeme
VUB AI-lab
Consensus
• Colour categories have a focal point and an
extent with fuzzy boundaries.
• Colour categories can be named.
• Different languages use different colour words.
• Colour categorisation aids our visual
perception.
• Mechanism of human colour perception…
Tony Belpaeme
VUB AI-lab
Human colour perception
• Human retina contains three types of chromatic
photoreceptors
• Combining the reaction of these three types provides
chromatic discrimination.
• From trichromacy to opponent channel processing
– Psychologically humans react in an opponent fashion to
colours.
Tony Belpaeme
VUB AI-lab
Controversies
• Are colour categories innate or learned?
• Shared within a language community?
• Shared between different cultures?
• If learned,
– What constraints are there on learning?
– Can learning explain sharedness?
• If culturally learned, does language have an
influence on colour categorisation?
Tony Belpaeme
VUB AI-lab
Support for universalism
• For example
– Berlin and Kay (1969).
– Rosch (1971, 1972).
Tony Belpaeme
VUB AI-lab
Berlin & Kay (1969)
• Experiment to identify colour categories in
different cultures through their linguistic
coding.
– Identified basic colour terms (BCT) of language.
– Asked subjects to point out the focus and extent of each
BCT.
Tony Belpaeme
VUB AI-lab
Berlin and Kay, results
Tony Belpaeme
VUB AI-lab
Rosch (1971, 1972)
• Experiments with Dugum
Dani tribe
– To demonstrate that colour
categories are not under the
influence of language.
– All confirmed that
categories were shared (and
thus innate) and not
influenced by language.
Tony Belpaeme
VUB AI-lab
Support for relativism
• Brown and Lenneberg (1954)
– Positive correlation between ‘codability’ of colour
terms and memorising colours.
• Davidoff et al. (1999)
– Reimplemented Rosch’s experiments.
– Unable to confirm Rosch, but instead support for
relativism.
• From 1990s
– Critical evaluation of 20 years universalism (Lucy,
Saunders & van Brakel).
– Evidence from subjects with anomalous colour vision
(Webster et al., 2000).
Tony Belpaeme
VUB AI-lab
Summary
Position
Acquisition
Sharing
Universalism/
nativism
Genetic
expression during
development
Gene propagation
Empiricism
Individual learning
Similar environment,
ecology and
physiology
Culturalism
Social and cultural
learning
Similar environment,
ecology and
physiology with
cultural learning
Tony Belpaeme
VUB AI-lab
Four experiments
Language
With
Without
Colour categories
Evolved
Learned
Individual
learning
Cultural
learning
Genetic
evolution
Genetic
evolution under
linguistic pressure
• Goal
– Study positions through computer simulations.
– Advance claims based on these simulations.
Tony Belpaeme
VUB AI-lab
Experimental setup
• An individual is modelled by an agent
–
–
–
–
Perception
Categorisation
Lexicalisation
Communication
• Agents are placed in a population
Tony Belpaeme
VUB AI-lab
Overview of an agent
Agent
Internal
representation
Perception
Tony Belpaeme
VUB AI-lab
Categories
Categorisation
Word forms
Lexicalisation
Perception
• Stimuli are presented as spectral power
distributions
• Modelling chromatic perception
– A model is needed
– Suitable for modelling categories on
Tony Belpaeme
VUB AI-lab
Perception
• CIE L*a*b* space
–
–
–
–
Perceptually equidistant space.
Similarity function exists.
Straightforward computation.
Suitable for defining colour categories on (Lammens,
1994).
Tony Belpaeme
VUB AI-lab
Categorisation
• Define categories on an internal colour
representation.
• Requirements
–
–
–
–
–
–
Delimiting regions in representation space
Measure of membership
Fuzzy extent
Learnable
Adaptable
Mutable
• Several possible representations, but the
choice fell on ‘adaptive networks’
Tony Belpaeme
VUB AI-lab
Adaptive network
• An adaptive network is radial basis function
network which is adapted instead of trained.
• One adaptive network represents one category
• Properties
– Fulfils all requirements.
– Based on exemplars.
– Can represent non-convex and asymmetrical
category shapes.
– Can be used as an instantiation of prototype theory
(Rosch).
– Easy to analyse
– Speedy
Tony Belpaeme
VUB AI-lab
Adaptive network
1
(
)
2
1
Tony Belpaeme
VUB AI-lab
(
)
…
2
J
J
(
)
Lexicalisation
• A category can be associated with no, one or
more word forms
• The strength of the association between a
word form and category is represented by a
score.
c
s1
s2
f1
f2
sn
fn
Tony Belpaeme
VUB AI-lab
Adaptive models
• Learning without language
– Implemented as discrimination games.
• Learning with language
– Implemented as guessing games.
• Steels et al
Language
With
Without
Colour categories
Learned
Evolved
Tony Belpaeme
VUB AI-lab
Individual
learning
Cultural
learning
Genetic
evolution
Genetic
evolution under
linguistic pressure
Discrimination game
• Discrimination serves as a task to force the
acquisition of categories.
– Serves as pressure to create new categories and
adapt existing categories.
– Also used to evaluate the categorical repertoire
Tony Belpaeme
VUB AI-lab
DG scenario
• Create context and chose topic.
O = o1, K , oN
• Agent perceives context.
o1, K , oN
® s1, K , sN
• Agent finds closest matching category for each
percept.
" c Î C : yc (si ) £ yˆ (si )
• Is topic matched by a unique category?
count ( cs' 1 , K , cs' N , cs' t
Tony Belpaeme
VUB AI-lab
)=
1
DG dynamics
• If the discrimination game fails, this provides
opportunity to create new or adapt old
categories.
Tony Belpaeme
VUB AI-lab
Guessing game
• Two agents are selected for playing a GG.
• Serves as task to generate a categorical
repertoire and associated lexicalisations.
Tony Belpaeme
VUB AI-lab
Guessing game scenario
• Two agents are selected; one speaker, one
hearer.
• A context is presented to both agents, the
speaker knows the topic.
• The speaker finds a discriminating category c
for the topic.
• It conveys the associated word form f to the
hearer.
• The hearer interprets the word form, finds the
associated category c’ and points out the
topic.
opoint = arg max (yc (oi ))
Tony Belpaeme
VUB AI-lab
GG dynamics
Game can fail at many points
• Speaker
– No discriminating category.
– No associated word form.
• Hearer
– Does not know the word form.
– Fails to point out the topic.
• Opportunity to extend and adapt categories and lexicon.
Tony Belpaeme
VUB AI-lab
Evolutionary models
• Genetic evolution without language
– Fitness evaluated by playing discrimination games.
Language
With
Without
Colour categories
Learned
Evolved
Tony Belpaeme
VUB AI-lab
Individual
learning
Cultural
learning
Genetic
evolution
Genetic
evolution under
linguistic pressure
Genetic operator
• Agents are endowed with the ability to have a
categorical repertoire (!).
• Categories are genetically evolved, instead of
a ‘genetic code’.
• Asexual reproduction.
Tony Belpaeme
VUB AI-lab
Genetic operator
• Mutation
–
–
–
–
Adding a category
Removing a category
Extending a category
Restricting a category
• Fitness measure
– Discriminative success
Tony Belpaeme
VUB AI-lab
Results without communication
• Learning categories
• Genetic evolution of categories
Language
With
Without
Colour categories
Learned
Evolved
Tony Belpaeme
VUB AI-lab
Individual
learning
Cultural
learning
Genetic
evolution
Genetic
evolution under
linguistic pressure
Individual learning
• Discriminative success
average discriminative success
1
0.8
0.6
0.4
0.2
0
0
200
400
600
game
N=10, lOl=3, D=50
Tony Belpaeme
VUB AI-lab
800
1000
Individual learning
• Category variance
50
category variance
40
30
20
10
0
0
200
400
600
game
Tony Belpaeme
VUB AI-lab
800
1000
Individual learning
• Categories of two agents on Munsell chart
• There is no sharing across populations
Tony Belpaeme
VUB AI-lab
Genetic evolution
• Discriminative success
average discriminative success
1
0.8
0.6
0.4
0.2
0
0
50
100
generation
N=10, IOI=3, D=50
Tony Belpaeme
VUB AI-lab
150
200
Genetic evolution
• Category variance
40
category variance
35
30
25
20
15
10
5
0
0
50
100
generation
Tony Belpaeme
VUB AI-lab
150
200
Genetic evolution
• Categories of two agents on Munsell chart.
• There is no sharing across populations.
Tony Belpaeme
VUB AI-lab
Summary
• Without communication
– Both approaches attain a categorical repertoire
functional for discrimination.
– Individual learning leads to a certain amount of
sharing, but no 100% coherence.
– Genetic evolution leads to complete sharing.
– Both approaches do not arrive at sharing across
populations.
– Timescale different.
Tony Belpaeme
VUB AI-lab
Results with communication
• Cultural learning.
Language
With
Without
Colour categories
Learned
Evolved
Tony Belpaeme
VUB AI-lab
Individual
learning
Cultural
learning
Genetic
evolution
Genetic
evolution under
linguistic pressure
Cultural learning
average discriminative success
• Discriminative success
1
0.8
0.6
0.4
0.2
0
0
10000
20000
30000
game
N=10, IOI=3,D=50
Tony Belpaeme
VUB AI-lab
40000
50000
Cultural learning
• Communicative success
average communicative success
1
0.8
0.6
0.4
0.2
0
0
10000
20000
30000
game
Tony Belpaeme
VUB AI-lab
40000
50000
Cultural learning
• Category variance
5
4.5
category variance
4
3.5
3
2.5
2
1.5
1
0.5
0
0
10000
20000
30000
game
Tony Belpaeme
VUB AI-lab
40000
50000
Cultural learning
• Categories of two agents on Munsell chart.
• There is no sharing across populations.
Tony Belpaeme
VUB AI-lab
Influence of communication on coherence
25
category variance
20
15
ratio
10
Without language
5
With language
0
0
2000
4000
6000
8000
10000
game
Tony Belpaeme
VUB AI-lab
12000
14000
16000
18000
20000
Influence of communication on coherence
Individual learning
Cultural learning
60
40
40
20
20
b
b
60
0
0
-20
-20
-40
-40
60
60
40
80
20
-20
Tony Belpaeme
VUB AI-lab
20
60
0
-20
40
-40
80
20
60
0
a
40
40
-40
L
a
20
L
Discussion on cultural learning
• Communication forces sharing in a cultural learning
through positive feedback between category formation
and communication.
• Communication has a causal influence on category
formation.
• First learning categories, and then lexicalising does
allow communication.
• Communicative success never 100%. In accordance
with anthropological experiments (Stefflre et al, 1966).
• Nature of categories is stochastic. Not in accord with
Berlin and Kay (1969).
• Model possibly does not contain enough ecological and
biological constraints.
Tony Belpaeme
VUB AI-lab
Summary
• Computer simulations on the acquisition of
colour categories.
• Extreme positions to allow a clear discussion.
• Both cultural learning and genetic evolution
seem to be good candidates for explaining
sharedness.
• Results and recent literature lend support for
culturalism.
Tony Belpaeme
VUB AI-lab
http://arti.vub.ac.be/~tony
Tony Belpaeme
VUB AI-lab
Tony Belpaeme
VUB AI-lab
Critical notes
• A computer simulation requires assumptions
and models.
Though results confirm the choices made, the assumptions
might be wrong.
• Weak ecological and biological constraints.
Stronger constraints might explain phenomena
now unaccounted for.
• Colour has been taken in isolation.
Tony Belpaeme
VUB AI-lab
Contributions
• Provide food for thought for disciplines other than AI.
• Formalisation of an interdisciplinary and often rhetoric
debate.
• Computer simulations of real world phenomena.
• Simulations with continuous meaning representation.
• A computational representation of natural categories.
Tony Belpaeme
VUB AI-lab
Artificial intelligence
Two kinds of AI
Constructing intelligence
Understanding intelligence
Building artefacts which
display adaptive or even
intelligent behaviour.
Studying complex
behaviour through
constructing artificial
systems.
Tony Belpaeme
VUB AI-lab
Situating the research
• The origins and evolution of language
– Humans are the only species mastering complex language.
– Humans possess complex cognitive abilities.
– Language might be the key to intelligence.
Tony Belpaeme
VUB AI-lab
The origins and evolution of
language
• Different lines of attack
– Linguistics
– Ethology
– Anthropology
– Artificial intelligence.
Tony Belpaeme
VUB AI-lab
The origins and evolution of
language
• Computers as a tool for investigating linguistic
phenomena
– Uses models and simulations.
– Allows investigation of mechanisms difficult or
impossible to study by other disciplines.
– Allows investigation of large parameter spaces.
– Provides no definite answers, but theories which are
referred back to observational disciplines.
Tony Belpaeme
VUB AI-lab
Various evidence for universalism
• Opponent neural response to chromatic stimuli
– Explains basic colour categories (Kay & McDaniel,
1978).
• Research on infants
– Infants possess colour categories for fundamental
colours (Bornstein et al., 1976).
Tony Belpaeme
VUB AI-lab
GG scenario
• Two agents are selected; one speaker, one
hearer.
• A context is presented to both agents, the
speaker knows the topic.
• The speaker finds a discriminating category c
for the topic.
• It conveys the associated word form f to the
hearer.
• The
hearer interprets the word form, finds the
opoint = arg max (yc (oi ))
associated category c’ and points out the
topic.
Tony Belpaeme
VUB AI-lab
Guessing game
Initialise the game
speaker
Tony Belpaeme
VUB AI-lab
hearer
Guessing game
Speaker discriminates topic
L
a
b
Tony Belpaeme
VUB AI-lab
Guessing game
Speaker finds word form associated with
category
=red
Tony Belpaeme
VUB AI-lab
Guessing game
Speaker conveys word form
red
Tony Belpaeme
VUB AI-lab
Guessing game
Hearer interprets word form
“Red”? Do I know this
form? If so, is it
uniquely related to a
stimulus?
Tony Belpaeme
VUB AI-lab
Guessing game
Hearer non-verbally points at topic
Tony Belpaeme
VUB AI-lab
Chromatic input
• Spectral
power
distributions
of actual chips
– Presented in
aperture
mode.
– Constant
adaptation
state.
– No
commitment
to any specific
device.
Tony Belpaeme
VUB AI-lab
Individual learning
• Changing environment
1
10
9
0.8
8
7
0.6
6
number of categories
5
0.4
4
3
0.2
2
1
0
0
10
20
30
40
50
game
Tony Belpaeme
VUB AI-lab
60
70
80
90
0
100
average number of categories
average discriminative success
DS
Genetic evolution
• Changing environment
12
DS
10
0.8
8
0.6
number of categories
0.4
4
0.2
2
0
0
20
40
60
generation
Tony Belpaeme
VUB AI-lab
6
80
0
100
number of categories
average discriminative success
1
Berlin and Kay, results
• Evolutionary order of basic colour terms.
épurple ù
ê
ú
green
ê
éwhit e ù
é
ù
pink ú
ê
ú
ê
ú < [red ] < ê
ú < [blue ] < [brown ] < ê
ú
êblack ú
ê
ú
yellow
orange
ú
ê
ú
ë
û
ëê
û
ê grey ú
êë
ú
û
• A language has at most 11 BCTs.
• Basic colour categories are genetically
determined.
Tony Belpaeme
VUB AI-lab
Cultural learning
• Number of categories
number of categories
14
12
10
8
6
4
2
0
0
10000
20000
30000
game
Tony Belpaeme
VUB AI-lab
40000
50000
Individual learning
• Number of categories
average number of categories
12
10
8
6
4
2
0
0
200
400
600
game
Tony Belpaeme
VUB AI-lab
800
1000
Genetic evolution
• Number of categories
number of categories
14
12
10
8
6
4
2
0
0
50
100
generation
Tony Belpaeme
VUB AI-lab
150
200
Genetic evolution with communication
• Number of categories
number of categories
14
12
10
8
6
4
2
0
0
50
100
150
200
generation
Tony Belpaeme
VUB AI-lab
250
300
350
400
Genetic evolution with communication
• Discriminative success
average discriminative success
1
0.8
0.6
0.4
0.2
0
0
50
100
150
200
generation
N=20, IOI=3, D=50
Tony Belpaeme
VUB AI-lab
250
300
350
400
Genetic evolution with communication
• Communicative success
communicative success
1
0.8
0.6
0.4
0.2
0
0
50
100
150
200
generation
Tony Belpaeme
VUB AI-lab
250
300
350
400
Genetic evolution with communication
• Category variance
10
9
category variance
8
7
6
5
4
3
2
1
0
0
50
100
150
200
generation
Tony Belpaeme
VUB AI-lab
250
300
350
400
Genetic evolution with communication
• Categories of two agents on Munsell chart.
• There is no sharing across populations.
Tony Belpaeme
VUB AI-lab
Discussion on genetic evolution with
communication
• Categories still evolve under communicative pressure.
• Sharedness within population arises through
propagation of genetic material.
• Not shared cross-culturally.
• Time-scale is radically different from cultural learning.
• Again, model possibly does not contain enough
ecological and biological constraints.
Tony Belpaeme
VUB AI-lab
Summary
• Learning with communication
– Both approaches attain a categorical repertoire and
lexicon.
– Both arrive at shared categories in the population.
– Both do not arrive at shared categories across
populations.
– No human-like categories.
Tony Belpaeme
VUB AI-lab