Layer 2/3 Is A Neural Multigrid: Information Maximization In A Local Network Explains V1 Hypercolumn Formation

One of the most striking feature of primate V1 is a topographic ordering of receptive fields. Previous models of topographic map formation (e.g., [1]) pose topography as a constraint on receptive-field formation. Optimization techniques maximize information between visual inputs and outputs [2], and produce receptive fields strikingly similar to those in V1, but not topography. Neural network implementations of infomax [3] require fully connected lateral networks, which the cortex does not possess. While the density of lateral connections in layer 2/3 surpasses all other connection classes [4], a local, patchy connectivity pattern is still observed [5]. A model of infomax in 2/3 must therefore explain how this optimization can be implemented in a local network. We hypothesized that topography might emerge naturally from such a model.