Neural network approaches to nonlinear blind source separation

Abstract

In this paper, several recently proposed neural network approaches to nonlinear blind signal separation (BSS) are reviewed. Of great interest, popular multilayer perceptron (MLP), radial basis function (RBF) and polynomial neural networks are the focus of the paper. In order to uniquely extract the original source signals from only nonlinearly mixed observations, some forms of constrains are always imposed on the neural networks. Three structurally constrained nonlinear independent component analysis mixing models are presented, followed by the discussion on additional signal constraints to the original cost function stemmed from the Kullback-Leibler Divergence.