A Novel FastICA Method for the Reference-based Contrast Functions

Abstract

This paper deals with the efficient optimization problem of Cumulant-based contrast criteria in the Blind Source Separation (BSS) framework, in which sources are retrieved by maximizing the Kurtosis contrast function. Combined with the recently proposed reference-based contrast schemes, a new fast fixed-point (FastICA) algorithm is proposed for the case of linear and instantaneous mixture. Due to its quadratic dependence on the number of searched parameters, the main advantage of this new method consists in the significant decrement of computational speed, which is particularly striking with large number of samples. The method is essentially similar to the classical algorithm based on the Kurtosis contrast function, but differs in the fact that the reference-based idea is utilized. The validity of this new method was demonstrated by simulations.