Spatiotemporal receptive fields maximizing temporal coherence in natural image sequences

Abstract

The relationship between the structure and functionality of the visual system and the properties of natural visual stimuli is an active research topic in computational visual neuroscience. It has previously been shown that maximization of temporal coherence of activity levels is one computational principle which leads to the emergence of simple-cell-like spatial receptive fields from natural image sequences. In this paper we extend previous results by examining the case of spatiotemporal receptive fields. We show that application of the same principle of temporal coherence in the spatiotemporal case yields receptive fields which are not only localized, oriented and multiscale, as in the spatial case, but also share some important temporal characteristics with spatiotemporal simple-cell receptive fields. Quantitative measurements of the properties of the resulting receptive fields are also provided, and these are compared against similar results obtained with independent component analysis.