Pattern Recognition Using Neural Networks. Comparison to the Nearest Neighbour Rule

Abstract The purpose of this paper is to compare two pattern recognition methods : Neural Networks and the Nearest Neighbour (NN) rule. As we shall see, a neural network approach corresponds to an approximation of the decision function. First of all, we present the neural network approach and the properties of the error function (which must be minimized) as a function of the network parameters, the a priori probabilities and the conditional probability density parameters. Considerations about the convergence of a classical steepest descent algorithm are presented, which lead to variable step method application. Since the learning “quality” is related to the number of cells in the network, we introduce a criterion for determining the optimum number of cells. Examples corresponding to discrimination beetwen two clusters in R 2 are shown and the results are compared to those obtained by the NN rule. Both methods will be applied to automatic analysis of human sleep. Our results will be compared to the visual classification achieved by an expert.