Experiments in Adaptive Pattern Recognition

Abstract

The purpose of this paper and the experiments which it describes has been to supply data concerning the power of some Perceptron-like adaptive pattern recognition systems using linear discriminate functions. Three problems have been presented to such a machine: hand-print classification, blood-cell sorting and target identification in gray-scale aerial photographs. Performance of decision functions utilizing corrective training were compared with that obtained by a simple form of Bayes' weighting. In general, the technique of corrective training was found to yield markedly superior results over the training sequence, but the ability to generalize or recognize samples not included in the training sequence was found to be about the same for the two techniques. Analysis of the experimental data permitted a quantitative evaluation of the effects of statistical dependence in the system together with a prediction of terminal error rates for the condition in which the number of A units is made infinitely large.