Performance and the Generalisation Error

Abstract

Generalisation (test set) and empirical (training-set) classification errors are meaningful characteristics of any pattern classification system. Generally, one needs to know both these error rates and their relationship with the training-set sizes, the number of features, and the type of the classification rule. This knowledge can help one to choose a classifier of the proper complexity, with an optimal number of features, and to determine a sufficient number of training vectors. While training the non-linear SLP, one initially begins with the Euclidean distance classifier and then moves dynamically towards six increasingly complex statistical classifiers. Therefore, utilisation of theoretical generalisation error results obtained for these seven statistical classifiers becomes a guide for analysing the small sample properties of neural net generated classification algorithms.