A RECURSIVE BAYESIAN APPROACH TO PATTERN RECOGNITION

The pattern classification problem is stated in terms of an ideal system and a model system. The ideal system gives the true classification of each input pattern while the model system gives an estimate of the probable classification of each pattern. The estimate is produced by the model as a parametrically defined function of the input pattern. The problem is to find a training algorithm which determines the values of the parameters, initially unspecified, from a sequence of inputs and outputs of the ideal and model systems and, thereby, determines the characteristics of the model. A general theoretical solution to this problem is given by a recursive Bayesian estimation procedure. Individual implementations may be derived by approximations which trade accuracy for computational ease. Assumptions of linearity and normality lead to a perceptron-like algorithm. Several non-linear and non-normal cases are also considered. © 1968.