Adaptive ICA with order statistics in multidimensional scenarios

Abstract

In this paper w e propose an alternativ e statistical Gaussianit y measure whose optimization provides the extraction of one non- gaussian independent component at each stage of an ICA procedure; this measure is based on the Cumulative Density Function (cdf) instead of traditional distribution distances over Probability Density Functions (pdf's). Additionally, a novel m ultistage-deflation algorithm is proposed in order to perform ICA in multidimensional scenarios very efficiently; although this approach can be applied to any multistage ICA method, we have developed it to speed up our ICA procedure based on Order Statis- tics (OS). The algorithm consists on a gradien tlearning rule plus an orthonormalization projection technique that decreases the vector space dimension progressively 1. © Springer-Verlag Berlin Heidelberg 2001.