Learning efficient internal representations from natural image collections

Abstract

Learning in sensory systems takes place after a repeated exposure to the incoming signals and many ideas based on information theoretical principles have been proposed to explain the synaptic adaptation which improves the coding capabilities of sensory areas. In this paper we want to emphasize that a simple, natural learning rule can be derived from a careful treatment of image redundancies. The learning rule is used to split images into independent components which connect different resolution levels, in a nonlinear way. The result shows the biological plausibility of this coding strategy not only in the visual pathway but also in other sensory modalities.