Training Set Selection in Neural Network Applications

Abstract

In some applications of ANNs to classification problems, training can be inefficient simply because of the high volume of data available for training purposes. The question arises whether it is necessary to use all the data in order adequately to approximate the decision boundaries.

It is intuitively obvious that points near to the boundaries are likely to have more influence over the network weights than those which are far away. Of course, as the boundaries are initially unknown, the status of any particular data point is also unknown. Here we describe an approach which uses an initial crude estimate of the decision boundaries to select appropriate training data in the case of the Multi-Layer Perceptron, followed by a phased addition of points to the training set. We compare this approach with the standard method on both artificial and real data sets, and report results which demonstrates the potential for improved performance in terms of both efficiency and reliability.