Organic Convolution Model of Ventral Visual Path Reproduces the Fine Structure of Shape Tuning in Area V4

Abstract

This modeling study proposes a novel theory for how V4 neurons embody selectivity to the varying degrees of curvature in diverse receptive field arrangements reported in previous recording studies. The model shows that a simple, unsupervised approach to curvature selectivity can explain a wide variety of past observations: V1 aggregates points into selectivity for lines and edges at a variety of rotations; V2 aggregates orientated segments into selectivity for corners at a variety of angles and rotations; V4 aggregates corners into curvature selectivity in a variety of degrees, rotations and positions. The model is implemented in 900,000 integrate and fire units in standard cortical micro-circuits obeying Dales law. The unit model and spike timing and transmission use realistic biophysical parameters. A novel method for simulating large numbers of integrate and fire units with tensor programming and GPU hardware is employed: By combining convolution with the time course of post-synaptic potentials, computation occurs in a feedforward cascade of loosely synchronized spikes. The model has relatively few parameters which are tuned by stochastic search with manual fine tuning. A synthesis of hierarchical and convolutional network models, this study adds novel elements to both: In comparison to previous hierarchical models there is a novel V4 mechanism and a biologically realistic computation based on single spikes using tensor convolution as the simulation engine. In comparison to previous convolution models there is a much higher degree of biological realism and a novel unsupervised approach to connection formation.