A Maximum Likelihood Approach to Nonlinear Blind Source Separation

In the basic signal model of blind source separation (BSS), an unknown linear mixing process is assumed. While this ensures under mild conditions a suuciently unique solution, it is desirable to extend the problem to nonlinear mixtures. Unfortunately the nonlinear case is much more diicult to handle, and brings serious indeterminacies to the solutions in the general case. In this paper we propose a new maximum likelihood approach to the nonlinear BSS problem. It is assumed that the source densities are known and that the mixing mapping is regularized. By nding a regular separating mapping which maximizes the likelihood, we show experimentally that the sources can often be separated.