Abstract
The response properties of simple and complex cells in V1 have been studied physiologically, and many models have been proposed. Note that a model must have adjustable parameters and be flexible enough to capture the diversity of response profiles of V1 neurons. On the other hand, the model must be rigid enough to be falsifiable. Using a combination of simulation experiments and mathematical analyses, we show how to satisfy these complementary conditions for the popular divisive normalization model (Heeger, 1992, Vis. Neurosci.; Carandini & Heeger, 2012, Nature Reviews Neurosci) for V1 neurons. A Matlab implementation of the model that can take static grayscale images as inputs was applied systematically to a battery of visual stimuli used in dozens of published physiological studies. We found that, with a small set of parameters consistent with empirical measurements, the model can account for over 25 phenomena in single-cell recordings of simple and complex cells in V1. Mathematical analysis showed that a parameter representing the maintained discharge (baseline firing rate) of the model neuron plays a critical role in three physiological phenomena: (A) the existence of inhibitory regions in the receptive fields of simple cells in V1, (B) the super-saturation effect in the contrast sensitivity curves, and (C) the narrowing/widening of the spatial-frequency tuning curves when the stimulus contrast decreases. Importantly, the association with a single parameter makes these three phenomena interdependent. This predicts that simple cells in V1 can be categorized in the following two mutually exclusive types: One type of cell shows phenomena A, B, and widening (C); the other shows not-A, not-B, and narrowing (C). This prediction of interdependence is potentially falsifiable. The simulation experiments and mathematical analyses show that the model satisfies the complementary desiderata of flexibility and rigidity.
Meeting abstract presented at VSS 2016