Taxonomy of Pattern Classification Algorithms

Abstract

Two or three hundred different pattern classification algorithms have been suggested in literature during the last 50 years. The main objective of this chapter is to review a selection of known statistical algorithms that can be obtained or improved by training ANN-based classification systems. The first selection contains seven statistical algorithms that can be obtained while training linear and non-linear single layer perceptrons and the second selection contains algorithms that can be approached in ANN training after deriving new non-linear features from the original ones. Particular attention is given to methods which can be used to structure the covariance matrices and describe them by a small number of parameters. This approach is not very popular in statistical pattern recognition, however, together with utilisation of neural networks, it becomes a powerful tool to solve problems in small training-set situations.