Emergence of strategic cone weighting from efficient coding of spatiochromatic natural images.

We develop an efficient coding model to address how a population of retinal ganglion cells (RGCs) can optimally combine signals from the retinal cone mosaic to maximize information transfer through the optic nerve. The model takes into account the redundancies inherent in color natural images and predicts how they should be reduced in order to make the best use of channel capacity in the optic nerve, given metabolic constraints, wiring constraints, and input and channel noise. RGCs are modeled as a set of linear-nonlinear neurons whose instantaneous firing rate is computed via a weighted sum of cone responses from a simulated L- and M-cone mosaic, followed by a rectifying nonlinearity. When adapted to a set of calibrated color natural images so as to maximize mutual information between the retinal image and RGC outputs, the learned weights exhibit a circularly symmetric, center-surround structure, and the population of RGCs tile visual space via ON- and OFF-mosaics, in line with previous studies that use only luminance variations in natural scenes. Over a range of cone-to-neuron ratios, the model RGCs strategically weight cones of a particular spectral type to construct a stronger form of L-M cone-opponency than would be obtained with purely random sampling, implying that such a specific arrangement increases information transfer through the optic nerve. Additionally, we find that the degree of cone-type-specific adaptation varies with the amount of noise in the cone activations, with less noise leading to more specific adaptation. The results of this study point to the benefits of strategic cone weighting for maximizing information transfer for spatiochromatic natural scenes.