Abstract
The optimal representation of a signal is determined by the task and the dominant noise. For high-level colour perception, the purpose of colour is to inform the observer about the world, and the dominant noise is due to failure of colour memory (Baddeley & Attewell, 2009). We estimated the properties of high-level colour perception by testing colour memory in CIE1931 chromaticity space. We observed categorical biases in colour memory across hue and saturation dimensions, where colour memory was biased towards category foci, corresponding to six basic colour terms: red, blue, green, pink, orange and grey (Berlin & Kay, 1969). We propose that these biases are due to a non-uniform prior over colours, which originates from the distribution of colours across objects in the environment. To identify the form of this prior, we trained a deep neural network to identify objects using only object colour. A single pixel was sampled from images of objects in ImageNet, and the model learnt to predict the probability of objects for a given pixel colour. We measured the amount of information provided by colour about objects across CIE1931 chromaticity space. Five colour categories were observed in the information geometry corresponding to basic colour terms, where category foci were more informative about objects than category boundaries. We replicated these results using the OpenImages V6 dataset, which produced a very similar categorical structure. The geometry of high-level colour perception reflected the information geometry of object colour space: colour memory was biased towards category foci which were more informative about objects, and away from category boundaries which were less informative about objects. These findings support the theory that the colour statistics of our environment form the basis of a non-uniform prior which directs perceptual processes towards the most informative colours, and explains the emergence of basic colour terms.