Numerical Aspects of Machine Learning in Artificial Neural Networks

Abstract

In previous papers e.g. [5] the effect on the learning properties of filtering or other preprocessing of input data to networks was considered. A strategy for adaptive filtering based directly on this analysis will be presented. We focus in Section 2 on linear networks and the delta rule since this simple case permits the approach to be easily tested. Numerical experiments on some simple problems show that the method does indeed enhance the performance of the epoch or off line method considerably. In Section 3, we discuss briefly the extension to non-linear networks and in particuar to backpropagation. The algorithm in its simple form is, however, less successful and current research focuses on a practicable extension to non-linear networks.