Multiresolution and compact support in neural network simulations
Toufic I. Boubez · 1993
Neurocomputing and the Neural Networks paradigm, especially the feed-forward (FFNN) model, have recently emerged as popular computing tools for the supervised learning of classifications and function approximations. The task typically requires the network to induce an input-output map from a set of labeled examples in an attempt to correctly classify input features that are not part of the original training set. The current model, however, suffers several shortcomings. First, the number of units that are necessary in the hidden layer(s) cannot be easily estimated before attempting to solve a problem, and is generally dependent on the problem structure, resulting in a long process of trial and error. Another problem lies with the back-propagation training method, which is a variant of gradient descent, and therefore susceptible to local minima, flatspots and numerical instabilities. Finally, the hyperplane classification inherent in most network models can be limiting and lacks a solid theoretical framework.