Neural Blind Source Separation using Second Order Autoassociators with Time-Delay.

Author: Konstantinos I. Diamantaras (Aristotelian University of Thessaloniki)

Abstract: The blind separation of sources is usually based on some independence assumption about the sources and higher-order statistics are used to achieve separation. Recently it has been shown that second-order statistics are also capable of separating sources provided that the source signals are pairwise uncorrelated and are not white in the time domain. If for example, the original sources are continuous-time, bandlimited signals non-whiteness in time can be achieved by oversampling. In this talk I propose a modification of a Hebbian type rule originally proposed for computing the cross-correlation SVD of a pair of signals and we show how this rule can be applied in the adaptive blind separation of uncorrelated sources. The modified model turns into a time-delay autoassociator of the input signal and closely resembles Oja's single-unit rule for PCA. There are three major advantages in unsing the second order approach: (a) the strong independence assumption about the source signals is reduced to the weak uncorrelatedness assumption, (b) the constraint that at most one signal can be Gaussian is removed, and (c) the Higher-order statistical optimization methods are replaced with simpler and better understood second order methods.