On FastICA algorithms and some generalisations

Abstract

The FastICA algorithm, a classical method for solving the one-unit linear ICA problem, and its generalisations are studied. Two interpretations of FastICA are provided, a scalar shifted algorithm and an approximate Newton method. Based on these two interpretations, two natural generalisations of FastICA on a full matrix are proposed to solve the parallel linear ICA problem. Specifically, these are a matrix shifted parallel ICA method and an approximate Newton-like parallel ICA method. © 2011 Springer Science+Business Media B.V.