A performance analysis of the multilayer perceptron in limited training data set conditions

This paper is dedicated to the experimental analysis of a limited training data set influence on the performance of multilayer perceptron as a nonparametric pattern classifier. The analysis is done in accordance to the following two criteria: the classification error and the sensitivity of the multilayer perceptron to the increasing number of features in limited training data set conditions. The experiments involve classifying both the synthesized data with known Bayes errors and the real data with unknown Bayes errors. The performances of the multilayer perceptron are analyzed with comparison to: the corresponding Bayes error estimates obtained by a very reliable estimation procedure based on the k-NN approach, the quadratic classifier, and the 3-nearest neighbor "volumetric" classifier.