Abstract
It has been proposed that one goal of visual processing is to achieve an efficient representation by reducing redundancy in the sensory inputs. In the case of trichromatic color vision of primates, redundancy reduction by decorrelation with orthogonal basis functions, such as Principal Component Analysis, leads to color-opponency as found in the retina and lateral geniculate nucleus. However, in visual cortex, color selectivities are not organized around orthogonal color space axes. To investigate the efficiency of non-orthogonal codes for color, we analyzed estimated human cone responses to natural color scenes. We used Independent Component Analysis to find linear transformations of the spatiochromatic data such that the outputs are statistically as independent as possible, i.e. least redundant. The resulting representations yielded highly sparse outputs and achieved higher coding efficiency than orthogonal codes. Color selectivities clustered around non-orthogonal axes in color space. In particular, one axis corresponded to the peak of the distribution of color selectivities in primary visual cortex. This suggests that the organization of color selectivities in visual cortex may reflect an adaptation to the distribution of color signals in natural scenes. The specific directions of the opponency axes depended not only on the distribution of natural spectra, but also on the shapes and overlaps of the cone spectral sensitivities. However, opponency emerged even with hypothetical cone spectral sensitivities that did not overlap. This indicates that color opponency is an efficient coding principle for natural spectra, regardless of the shapes and overlaps of the photoreceptor sensitivities.