Conference on Cognitive Computational Neuroscience

Convolutional neural networks exhibit organizational principles of early visual areas. However, they still lack several important characteristics such as cortical magnification, eccentricity-dependence of receptive field sizes and spatial frequency tuning, as well as radial bias. We suggest that these properties arise from the non-uniform sampling of external space that results from the distribution of ganglion cells in the retina. To test this conjecture, we introduce the retinal sampling layer (RSL) which resamples images according to retinal ganglion cell density, before feeding them into the CORnet-Z architecture. Training the network on natural images resulted in it developing all of the hypothesized characteristics.