Unsupervised learning in constrained linear networks

Constrained linear architectures which learn in unsupervised mode according to Hebb's rule to minimize the output energy are analyzed. Under which conditions such networks act as decorrelating (square-root) filters is investigated. An algorithm is proposed which is almost optimum since the inputs of each filter (which are also the outputs of the net) become more and more decorrelated as the algorithm progresses. The search essentially approaches Newton's algorithm. The results of a simulation are shown.<
>