Post-nonlinear Blind Source Separation Using Neural Networks with Sandwiched Structure

Abstract

This paper proposes a novel algorithm based on informax for post-nonlinear blind source separation. The demixing system culminates to a neural network with sandwiched structure. The corresponding parameter learning algorithm for the proposed network is presented through maximizing the joint output entropy of the networks, which is equivalent to minimizing the mutual information between the output signals in this algorithm, whereas need not to know the marginal probabilistic density function (PDF) of the outputs as in minimizing the mutual information. The experimental results about separating post-nonlinear mixture stimulant signals and real speech signals show that our proposed method is efficient and effective.