MISEP-an ICA method for linear and nonlinear mixtures, based on mutual information

Presents MISEP, an independent components analysis method for linear and nonlinear mixtures. The method is based on the minimization of the mutual information of the estimated components, and can be seen as an extension of the well known INFOMAX method in two aspects: (1) allowing the analysis of nonlinear mixtures, and (2) automatically estimating the optimal nonlinearities to be used at the outputs during learning. The resulting ICA method consists of the training of a single specialized multilayer perceptron, optimized according to a single objective function: the output entropy. Some examples of the application of the method are given.