Blind source separation using principal component neural networks

Abstract

Blind source separation (BSS)is approached from the second order statistics point of view. In particular, it is shown that temporal filtering by an arbitrary filter combined with PCA leads to the solution of the problem provided that the sources are colored and have different spectra. This result is demonstrated by applying a neural PCA model such as APEX to BSS problems with artificially created, randomly colored data.