Modeling the emergence of perceptual color space in the primary visual cortex

Abstract

Humans’ perceptual experience of color is very different from what one might expect, given the light reaching the eye. Identical patterns of light are often perceived as different colors, and different patterns of light are often perceived as the same color. Even more strikingly, our perceptual experience is that hues are arranged circularly (with red similar to violet), even though single-wavelength lights giving rise to perceptions of red and violet are at opposite ends of the wavelength spectrum. The goal of this thesis is to understand how perceptual color space arises in the brain, focusing on the arrangement of hue. To do this, we use computational modeling to integrate findings about light, physiology of the visual system, and color representation in the brain. Recent experimental work shows that alongside spatially contiguous orientation preference maps, macaque primary visual cortex (V1) represents color in isolated patches, and within those patches hue appears to be spatially organized according to perceptual color space. We construct a model of the early visual system that develops based on natural input, and we demonstrate that several factors interact to prevent this first model from developing a realistic representation of hue. We show these factors as independent dimensions and relate them to problems the brain must be overcoming in building a representation of perceptual color space: physiological and environmental variabilities to which the brain is relatively insensitive (surprisingly, given the importance of input in driving development). We subsequently show that a model with a certain position on each dimension develops a hue representation matching the range and spatial organization found in macaque V1—the first time a model has done so. We also show that the realistic results are part of a spectrum of possible results, indicating other organizations of color and orientation that could be found in animals, depending on physiological and environmental factors. Finally, by analyzing how the models work, we hypothesize that well-accepted biological mechanisms such as adaptation, typically omitted from models of both luminance and color processing, can allow the models to overcome these variabilities, as the brain does. These results help understand how V1 can develop a stable, consistent representation of color despite variabilities in the underlying physiology and input statistics. This in turn suggests how the brain can build useful, stable representations in general based on visual experience, despite irrelevant variabilities in input and physiology. The resulting models form a platform to investigate various adult color visual phenomena, as well as to predict results of rearing experiments.