Recursive generalized eigendecomposition for independent component analysis

Abstract

Independent component analysis is an important statistical tool in machine learning, pattern recognition, and signal processing. Most of these applications require on-line learning algorithms. Current on-line ICA algorithms use the stochastic gradient concept, drawbacks of which include difficulties in selecting the step size and generating suboptimal estimates. In this paper a recursive generalized eigendecomposition algorithm is proposed that tracks the optimal solution that one would obtain using all the data observed. © Springer-Verlag Berlin Heidelberg 2006.