A neural network approach to blind source separation

The problem of adapting linear multi-input-multi-output systems for unsupervised separation of linear mixtures of sources arises in a number of signal processing applications. In this paper we present a new single layer neural network in which information transfer maximization is equivalent to minimizing a cost function involving the well-known constant modulus criterion originally used in blind equalization. The proposed approach is able to separate sources with negative kurtosis as revealed by an analysis of the cost function stationary points. Two learning rules are presented to compute the optimum separating matrix. One of them turns out to be an equivariant algorithm whose convergence does not depend on the mixture matrix.