Adaptive separation of an unknown number of sources

Abstract

The problem of separation of mixed sources is addressed in this paper. To solve this problem, at least as many observations as sources are needed. In particular, the number of sources can be unknown. The separation system is a linear network updated with a stochastic descent algorithm to minimize some separation criterion. A first algorithm separates sources with positive kurtosises while a second one separates sources with negative kurtosises. For both, the performances are independent of the mixture. Besides, in the noisy case, when there are more sensors than sources, the additional outputs merely generate noise.