Low complexity functions for stationary independent component mixtures

Abstract

Obtaining a low complexity activation function and an online sub-block learning for non-gaussian mixtures are presented in this paper. The paper deals with independent component analysis with mutual information as a cost function. First, we propose a low complexity activation function for non-gaussian mixtures, and then an online sub-block learning for stationary mixture is introduced. The size of the sub-blocks is larger than the maximal frequency Fmaxof the principal component of the original signals. Experimental results proved that the proposed activation function and the online sub-block learning method are more efficient in terms of computational complexity as well as in terms of learning ability.