Efficient confidence bounds for RBF networks for sparse and high dimensional data

Abstract

Artificial Neural Networks have been used for function approximation and pattern recognition in a variety of domains. However, due to its empirical nature, it is difficult to derive an estimate of neural network's accuracy. There are in the literature a number of proposed methods to calculate a measure of confidence to the output of neural networks but in general these approaches require some strong assumptions which are rarely observed in real problems. This paper analyzes and extends the Validity Index Network, a model derived from radial basis function network that calculates the confidence of its outputs, we remove it restrictions in density calculations, specially in high dimensional input spaces, and improve the probability coverage of the prediction levels when the training data have variable density. © 2010 Springer-Verlag Berlin Heidelberg.