Nonlinear Blind Source Separation by Self-Organizing Maps

Abstract

In neural blind source separation most approaches have considered the linear source separation problem where the input data consist of unknown linear mixtures of unknown independent source signals. The solution is a linear transformation which makes the output vector components statistically independent. More generally we can consider nonlinear mixtures of sources. Then we can try to separate the sources by constructing mappings that make the components of the output vectors independent. We show that such a mapping can be approximately realized using self-organizing maps with rectangular map topology. We apply these mappings to the separation of nonlinear mixtures of sub-Gaussian sources.