An auditory system for efficient coding of natural sounds

Abstract

This article presents a model of the auditory pathway. The model consists of two parts; one is nonlinear transformation and the other is sparse coding to reduce the dependency involved in the transformed signal. The later part theoretically corresponds to noisy Independent Component Analysis. The two parts individually learns so as to maximize the entropy. The model can well reproduce a couple of biological phenomena observed in the auditory system. They are virtual pitch and masking effect. These results imply that the nonlinear transformation by single neurons and the transformation realized by neural populations play essential roles to obtain efficient information processing, i.e., coding, in the primary auditory system. This is consistent to the recent results in the primary visual system, which have introduced the notion of maximum entropy criterion and sparse-coding.