Transcript Innateness of colour categories is a red herring: insights from

Innateness of colour categories is a
red herring: insights from
computational modelling
Tony Belpaeme
Artificial Intelligence Lab
Vrije Universiteit Brussel
Colour categories
• The colour spectrum is continuous
• Still, we divide it into colour categories
• What are the origins of colour categories?
(Insights might be applicable to other perceptual
categories as well)
Importance for language
“… this may at first appear to be a comparatively
trivial example of some minor aspect of
language, but the implications for other aspects
of language evolution are truly staggering.”
(Deacon, 1997)
Universalism
• Berlin and Kay (1969) used naming experiments
to extract colour categories
Universalism
• This universal character has been hailed by
many and has been reconfirmed by some.
(among others Rosch-Heider, 1972; Kay and McDaniel, 1978; Durham, 1991; Shepard, 1992; Kay
and Regier, 2003)
Three positions
• Supposing we accept a certain universality of
colour categorisation, what mechanisms could
underlie this?
– Nativism:
genetic makeup.
– Empiricism: interaction with the environment.
– Culturalism: cultural interaction with others.
Nativism
• Colour categories are
directly or indirectly
genetically specified.
L
+
_
+
+
– Regularities in human early
visual perception,
especially the opponent
character of colour vision.
(Kay and McDaniel, 1978)
– Regularities in the neural
coding of the brain.
(Durham, 1991)
– Genetic coding of colour
categories.
(Shepard, 1992)
L
S
M
+
+
+
R-G
_
+
Y-B
Empiricism
• Our ecology contains a certain chromatic
structure which is reflected in our colour
categories.
• We extract colour categories by interacting
with our environment.
(e.g. Elman et al., 1996; Shepard, 1992; Yendrikhovskij, 2001)
• This all happens without the influence of
culture or language.
Culturalism
• Colour categories are culture-specific.
• They are learned with a strong causal influence
of language and propagate in a cultural
process.
(e.g. Whorf, 1954; Davidoff et al., 2001; Roberson, 2005;
Belpaeme and Steels)
Nativism, empiricism or culturalism?
• The discussion has been held on many different
fronts
–
–
–
–
–
–
Neurology.
Psychology.
Anthropology.
Linguistics.
Ophthalmology.
Philosophy.
• We will tackle the discussion from artificial
intelligence and computer modelling.
How can Artificial Intelligence help?
• Artificial Intelligence allows us to create
models of natural phenomena, of which we
then observe their behaviour.
• Different premises can be implemented in the
models, allowing us to get an insight into the
validity of the premises.
– E.g. traffic modelling.
Studying empiricism
• Procedure
– Collect chromatic data.
– Extract colour categories. For this we use a
clustering algorithm.
– Compare extracted categories with each other and
with human colour categories.
• If empiricism holds, we would expect a high
correlation between the extracted categories
and human categories.
Chromatic data
• Three data sets: natural, urban and random
Extracting categories
Categoriesfrom
fromnatural
urban data:
Categories
data:
Quantitative comparison
• 11 categories extracted from natural and urban
data
100
• Correlation with human colour categories
80
60
40
b
20
0
-20
-40
-60
-80
-100
-80
-60
-40
-20
0
20
a
40
60
80
100
Reflections on empiricism
• The claim that human colour categories are
specified by the distribution of chromatic
stimuli in the world is not supported by our
data.
• However, there does seem to be a twofold
influence by
– The structure of the perceptual colour space.
– The properties of perceptual categories.
Studying culturalism
• Procedure
– Take a population of simulated individuals that learn
colour categories and communicate about colour.
• If culturalism holds, we expect linguistic
interactions to cause sharing of colour
categories.
The simulations
• Agent-based simulations
– An agent is a simulated individual, with perception,
categorisation, lexicalisation and communication.
– Perception maps spectral power distribution onto an
internal colour space.
– Categorisation maps percepts onto categories,
categories have prototypical behaviour.
– Lexicalisation connects categories to words.
– Communication takes care of uttering word forms.
– The agents have no way to access the internal state
of other agents: there is no telepathy!
Results
• Colour categories of two agents
• Agents arrive at colour categories that are
“shared”.
Results (2)
40
40
35
35
30
40
b
20
0
25
Ratio
40
25
20
20
20
0
15
15
-20
-20
10
-40
30
60
b
category variance
60
10
-40
Without language
60
60
5
40
20
-20
0
-40
a
20
60
0
0
5
40
80
1000
20
402000
3000
L
5000
game
6000
7000
8000
a
80
With language
0
4000
ratio between cv without and with language
• Influence of linguistic interactions on categories.
60
0
-20
9000
-40
10000
20
40
L
• But as language is culture-specific, cultural
evolution cannot explain universalism.
Summary
• Empiricism is not a good candidate to explain
universalism
– There is not enough ecological pressure.
• Culturalism can explain the sharing of categories in a
culture, but not universalism.
• Nativism can explain universalism, but is to slow to
follow ecological changes.
– Also, recent neurophysiological and molecular studies point out
many differences in colour perception between individuals.
Conclusion
• A blend of all three positions is needed to explain
universalism.
• But language and culture play a crucial role as the
catalysts which binds the perceptual categories of
individuals.
• Read the full story at http://arti.vub.ac.be
Steels & Belpaeme (2005) Coordinating Perceptually
Grounded Categories through Language: A Case Study
for Colour. Behavioral and Brain Sciences. To appear.