Hands on demonstration

Modelling the evolution of language for modellers and non-modellersIJCAI 2005

1

Hands on demonstrationHands on demonstration

Nature of colour categories


2

What do we know?What do we know?• There is evidence that colour categories are

universal.– All cultures have colour categories that are

similar to RED, GREEN, BLUE, YELLOW, and so on.

• Three possible explanations– Genetically determined.– Culturally (and linguistically) determined.– Ecologically determined.

• In this demonstration we will take a closer look at the last explanation.


3

Why do we want to know?Why do we want to know?• Language is about communicating

concepts, we want to now how concepts are acquired.

• As a case study we take colour categories.

• “…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)


4

Hypothesis and assumptionsHypothesis and assumptions• Research question

– Does our ecology contain enough structure to specify colour categories the way they are?

• Hypothesis– Human ecologies contain enough structure to specify

human colour categories.

• Assumptions– No semantics, culture or language is involved.– Colour categories have a prototypical nature.– Colour categories are extracted from chromatic stimuli

in an unsupervised manner.– We choose a representation for our colours that is

psychophysically plausible.


5

What do we expect?What do we expect?• If the claim is true:

– Categories extracted from the real-world data should resemble human colour categories.

– Categories extracted from the random data should not resemble human colour categories.

– Categories extracted from real-world data should not resemble the ones from random data.


6

ToolsTools• A digital camera.• Matlab (a mathematical package).• SPSS (a statistics package).


7

MethodologyMethodology• Gather image collection from natural and urban

environments.

• Draw 25,000 random pixels from each collection. Construct random set as control.


8

MethodologyMethodology• Extract categories from the data

– This we do by unsupervised clustering (k-means clustering) as this does not violate our assumptions.

• Compare the categories to human colour categories– Sturges & Whitfield (1995) have recorded the 11

basic colour categories of American English-speaking informants.

– Quantitative and objective comparing happens through matching couples and calculating the correlation between clusters and human colour categories. We use Kendall’s Tau correlation for ranked and matched observations.


9

The colour stimuliThe colour stimuli

natural urban

random


10

Extracted categories versus human categoriesExtracted categories versus human categories

0 10 20 30 40 50 60 70 80 90 1000102030405060708090100

l for Sturges and Whitfield

l for cluster

s

NATURE vs Sturges and Whitfield (1995) in CIElab


11

DemonstrationDemonstration• A quick demonstration of a “light” version

of an experiment.


12

ResultsResults

0 10 20 30 40 50 60 70 80 90 1000102030405060708090100

l for Sturges and Whitfieldl for c

lusters


-60 -40 -20 0 20 40 60 80-60-40-20020406080

a for Sturges and Whitfield

a for cluster

s


-60 -40 -20 0 20 40 60 80 100-60-40-20020406080100

b for Sturges and Whitfield

b for cluster

s


0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

90

100


l for

clu

sters

URBAN vs Sturges and Whitfield (1995) in CIElab

-60 -40 -20 0 20 40 60 80-60-40-20020406080

a for Sturges and Whitfielda for c

lusters


-60 -40 -20 0 20 40 60 80 100-60-40-20020406080100

b for Sturges and Whitfield

b for cluster

s


0 10 20 30 40 50 60 70 80 90 1000102030405060708090100


l for cluster

s

RANDOM vs Sturges and Whitfield (1995) in CIElab

-60 -40 -20 0 20 40 60 80-80-60-40-20020406080

a for Sturges and Whitfield

a for cluster

s


-60 -40 -20 0 20 40 60 80 100-80-60-40-20020406080100

b for Sturges and Whitfieldb for c

lusters


l a b

nature

urban

random


13

CorrelationCorrelationCorrelations between lightness, colour axes, chroma and hue.

Correlation between random

distribution and human

categories is not lower than for a real-world

distribution

For two different colour appearance models (CIE L*a*b* and CIE L*u*v*).


14

ConclusionConclusion• We could not refute the null hypothesis.

– Clustering random colours produces categories that correlate equally well.

• Human ecologies have only a marginal influence on colour categories.

• What then does have an influence?– Psychophysical properties of colour perception.– The nature of categories (maximally distinct).– And possible culture and language (but no proof

in this experiment).


15

More on thisMore on this• Yendrikhovskij, S.N. (2001) Computing Color Categories

from Statistics of Natural Images. Journal of Imaging Science and Technology, 45(5):409-417.

• Belpaeme, T. & Bleys, J. (2004) Does structure in the environment influence our conceptualization? Proceedings of the Conference on the Evolution of Language 2004, Leipzig, Germany.

• Steels, L. & Belpaeme, T. (2005) Coordinating perceptually grounded categories through language: A case study for colour. Behavioral and Brain Sciences, 28(4). In press.

Hands on demonstration

Documents