Stiefel-Grassman flow (SGF) learning: further results

Abstract

The aim of this this paper is to present recent contributions to Stiefel-Grassman flow (SGF) learning algorithms, a new class of learning paradigms for neural layers which allow for orthonormal signal/data processing. SGF learning has been introduced by the present author in 1996 as a way of training linear neural layers dedicated to blind source separation. In the meantime, several contributions have appeared in the scientific literature concerning the same topic, thus the study of a general framework explaining the different results has become necessary. In previous papers we presented a learning theory which appeared general enough to encompass the existing approaches; in this paper the latest results found are reported and discussed and references are given to computer simulations performed in order to test the effectiveness of the algorithms.