A Linear Adaptive Neural Network for Extraction of Independent Components

摘要

In this paper we introduce a linear adaptive neural network for extracting all independent components contained in a linear mixture. Each neuron of the network is updated with the same extended local anti-Hebbian rule [7] and is capable of extracting one component. However the adequate initialisations are hard to perform. Therefore we add a second global term in the learning rule of each neuron that involves informations from the other neurons and forces them to extract different components. When there are at least as many observations as components, all the components are extracted, at least by one neuron.