Adaptive Blind Source Separation by Auto-Associative Neural Network with Pruning

摘要

This paper describes a non-information theoretic approach to the Blind source Separation (BSS) by using the conventional auto-associative neural network (AANN). There is no special computation explicitly intended for BSS, except for a pruning mechanism which automatically determines the number of the sources. The nonlinear hidden units that have survived the pruning would recover the source signals separately. A computer simulation study is mode with examples to show adaptability of the present approach. A mathematical discussion presented relates BSS with local minima of the identity transformation error in AANN, and indicates the importance of the non-linearity, as opposed to the case of PCA by AANN.