Neural Networks Approach to Blind Source Separation
via Adaptive Independent Component Analysis
An original self-organizing neural network models for solution of blind source separation problem via adaptive independent technique is proposed. The purpose of research consists in generalization of a wiled-known method of independent component analysis to adaptive neural networks model. We develop an algebraical approach to unsupervised learning rule of heterogeneous neural nets with adaptive flexible activation. In frame of this approach we develop new information adaptation filtering model with entropy-based learning rule for self-organizing neural network. The main feature of our model is a flexible non-linearity of neuron activation that can be tuned in accordance to a true signal distribution. This allows accurately separating informative components of vector signals with different types of probability distribution functions. The article focuses an extended independent component analysis algorithm for mixtures of arbitrary signals with non-gaussian distributions.