Constructive neural networks as estimators of bayesian discriminant functions

Abstract

Of crucial importance to the successful use of artificial neural networks for pattern classification problems is how the appropriate network size can be automatically determined. This issue is addressed by formulating the process as an automatic search in the space of functions that corresponds to a subclass of multilayer feedforward networks. Learning is thus a dynamic network construction process which involves adjusting both the network weights and the topology. Adding new hidden units corresponds to extracting higher-level features from the original input features for reducing the residual classification errors. It can be shown that the resultant network approximates a Bayesian classifier that implements the Bayesian decision rule for classification. The empirical results of several pattern classification experiments are also reported.