Pre-processing and feature extraction in radial basis functions networks

Although neural networks have been born in engineering field, they are actually receiving a lot of attention among the statisticians. As a matter of fact, neural networks can be viewed as computational models very similar to statistical models that can be applied on several types of real data set. In this respect, as it is done in statistics, an important step of feature extraction and data transformation should be added to the learning and prediction phases of a neural network. The aim of this paper is to show the influence of this phase, called pre-processing, in radial basis function networks, on a real data set. The success of these networks, with and without data pre-processing, is measured by the discrimination rule and the generalisation to unobserved pattern. The performances of radial basis functions networks are also compared with the results obtained by the discriminant analysis, on the same data set.