Blind separation with unknown number of sources based on auto-trimmed neural network

This paper focuses on blind source separation with an unknown number of sources, which is the case generally assumed in most practical applications. Several over-determined neural algorithms (more sensors m than sources n) have been proposed to solve the problems associated with these cases, but separating performance is often sacrificed in order to prevent divergence. The general natural gradient descent can be validly applied to determined algorithms (m=n) only. Therefore, to better solve the problems, an algorithm associating the feed-forward neural network and an auto-trimming technique is proposed. The learning process starts with an over-determined architecture, followed by two steps used in every iteration. First, the number of sources is estimated by using the stability discriminant function, next, the neural network gradually trims redundant nodes according to an instant estimation. Validity and performance of the proposed approaches are demonstrated with computer simulations on artificially synthesized signals and compared with the well-known algorithm proposed by Ye et al.