Maximizing the information transfer for adaptive unsupervised 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 applications in multiuser wireless communications, such as mobile telephony. In this paper we propose a new statistical criterion to adapt the separating system. It involves the well-known Godard criterion as part of it and is interpreted by information theory as the maximization of information transfer in a single layer nonlinear neural network. The proposed criterion is free from undesirable stationary points provided that the signals to be separated have negative kurtosises, which is the case in communications.