Unsupervised Learning of Cone Spectral Classes from Natural Images

The first step in the evolution of primate trichromatic color vision was the expression of a third cone class not present in ancestral mammals. This observation motivates a fundamental question about the evolution of any sensory system: how is it possible to detect and exploit the presence of a novel sensory class? We explore this question in the context of primate color vision. We present an unsupervised learning algorithm capable of both detecting the number of spectral cone classes in a retinal mosaic and learning the class of each cone using the inter-cone correlations obtained in response to natural image input. The algorithm's ability to classify cones is in broad agreement with experimental evidence about functional color vision for a wide range of mosaic parameters, including those characterizing dichromacy, typical trichromacy, anomalous trichromacy, and possible tetrachromacy.Author Summary: The human visual system encodes color by comparing the responses of three different kinds of photoreceptors: the long- (reddish), medium- (greenish), and short- (bluish) wavelength-sensitive cone cells. In order for the visual system to accurately represent the color of stimuli, it must (in effect) know the class of the cone that produced each response. The long- and medium-wavelength-sensitive cones, however, are virtually identical in every known way except that their responses to a given spectrum of light differ. Here, we simulate cones in a model human retina and show that by examining the correlation of the responses of cones to natural scenes, it is possible to determine both the number cone classes present in a retinal mosaic and to explicitly determine the class of each cone. These findings shed light on the computational mechanisms that may have enabled the evolution of human color vision, as well as on the more general question of whether and when it is possible for sensory systems to self-organize.