Neural network schemes for blind separation of sources from nonlinear mixtures

Most existing BSS algorithms are based on the ideal situation where the mixture is merely a linear transformation of the source signals and the demixer is simply a linear network. Nonlinear techniques are presented for instantaneous blind signal separation using an information theoretic approach combined with (nonlinear) neural networks. Firstly, we address the issue of modelling the mixture for both linear and nonlinear transformation of the source signals. Secondly, we derived the required algorithm to train the variable gradient multilayer perceptron (MLP) based on a Lie group. In the past, most demixers employed a fixed gradient. Finally, computer simulations are carried out to compare the performance of the linear and nonlinear demixer when the underlying mixture of the source signals is either linear or nonlinear.