Modified post-nonlinear ICA model for online neural discrimination

The nonlinear independent component analysis (NLICA) is an extension of the standard ICA model that does not restrict the mixing system to be linear. Different algorithms have been proposed to solve the NLICA problem, but, as the dimension of the problem increases, most of them present limitations such as poor accuracy and high computational cost. In this work, a novel structural model is proposed for the overdetermined NLICA problem (when there exist more sensors than sources), by adding a signal compaction block to the standard post-nonlinear (PNL) de-mixing model. The proposed methodology proves to be efficient in the feature extraction phase of a challenging high-dimensional online neural discrimination problem.