Neural net with two hidden layers for non-linear blind source separation

In this paper, we present an algorithm that minimizes the mutual information between the outputs of a multilayer perceptron (MLP). The neural network is then used as separating system in the nonlinear blind source separation problem. It has been reported that MLPs with one hidden layer are sufficient to achieve desirable performance. However, in some cases, we may prefer approximating nonlinear mappings by using networks with several hidden layers. For the sake of simplicity, the present paper is focused on MLPs with two hidden layers. The performance is illustrated by some experiments.