The distribution of retinal ganglion cells in primate visual systems portrays a densely distributed central region, with an incrementally decreasing cell density as the angle of visual eccentricity increases. This results in a non-uniform sampling of the retinal image that resembles a wheelbarrow distortion. We propose that this sampling gives rise to several organizational properties of the primate visual system, including cortical magnification, linear relationship between eccentricity and receptive field sizes, eccentricity-dependent drop-off in spatial-frequency preference, and radial bias. We test this hypothesis by training a convolutional neural network to classify the orientation of sine gratings and Gabor stimuli, resampled according to retinal ganglion cell distributions. Our simulations show that introducing this sampling step gives rise to the aforementioned organizational principles in convolutional layers while only minimally affecting their classification performance. This lends credence to the notion that the retinal ganglion cell distribution is an important factor for the emergence of these organizational principles in visual systems.

Primate visual cortex exhibits key organizational principles: Cortical magnification, eccentricity dependent receptive field size and spatial frequency tuning as well as radial bias. We provide compelling evidence that these principles arise from the interplay of the non-uniform distribution of retinal ganglion cells (RGCs), and a quasi-uniform convergence rate from the retina to the cortex. We show that convolutional neural networks (CNNs) outfitted with a retinal sampling layer, which resamples images according to retinal ganglion cell density, develop these organizational principles. Surprisingly, our results indicate that radial bias is spatial-frequency dependent and only manifests for high spatial frequencies. For low spatial frequencies, the bias shifts towards orthogonal orientations. These findings introduce a novel hypothesis about the origin of radial bias. Quasi-uniform convergence limits the range of spatial frequencies (in retinal space) that can be resolved, while retinal sampling determines the spatial frequency content throughout the retina.

No abstract