A conventional auto = associative neural network separates blind sources without adding intentional algorithms other than pruning

摘要

A conventional auto-associative neural network (AANN) is shown to have an intrinsic ability to solve the Blind Source Separation (BSS) problem without special computations explicitly intended for BSS, except for a pruning mechanism to deal with the usual case in which the number of the sources is unknown; each nonlinear hidden unit that has survived the pruning would recover one of the source signals. The feasibility of this non-information theoretic approach is shown by computer simulation for two- and three-source examples involving various p.d.f.'s for the independent sources. A mathematical analysis is made to discuss BSS in the context of local minima associated with the nonlinearity-induced error in the identity transformation by the AANN.