Numerically Robust Learning Algorithms for Feed Forward Neural Networks

Abstract

In recent years several neural networks learning algorithms have been developed by making use of the RLS recursion. These algorithms are based on the matrix inversion lemma and in some cases can be numerically ill-conditioned. For example the rounding errors can accumulate and cause errors to occur in both the estimated parameters and the covariance matrix. In this paper we present numerically robust learning algorithms based on the QR decomposition - a well known technique in the linear prediction theory.