ABSTRACT
Pattern classification consists in assigning entities, described by feature vectors, to pre-defined groups of patterns. When the statistical characteristics of the problem under consideration are perfectly known, minimal error probability can be achieved by means of the Bayes decision rule. In practice, however, a suboptimal classifier has to be constructed from training data. Several neural network approaches to this problem have been proposed. Nearest-neighbor models are based on assessing the similarity between the input pattern and a set of reference patterns with known classification. The regression approach consists in predicting category from pattern by minimizing a certain error criterion. In the finite sample case, the definition of the structural complexity of these models is shown to have considerable influence on classification error. Finally, a taxonomy of the main neural network and alternative techniques of pattern classification are presented.
