Method for modelling similarity function using neural network

Abstract

Given a set of objects (A, B, C, . . . ), each described by a set of attribute values, and given a classification of these objects into categories, a similarity function accounts well for this classification when only a small number of objects are not correctly classified. A method for modelling a similarity function using a neural network comprises the steps of: (a) inputting feature vectors to a raw input stage of a neural network respectively for object S in the given category, for other objects G in the same category being compared the object S, and for object B outside the given category; (b) coupling the raw inputs of feature vectors for S, G, and B to an input layer of the neural network performing respective set operations required for the similarity function so as to have a property of monotonicity; (c) coupling the input elements of the input layer to respective processing elements of an hidden layer of the neural network for computing similarity function results adaptively with different values of a coefficient w of the similarity function; (d) coupling the processing elements of the hidden layer to respective output elements of an output layer of the neural network for providing respective outputs of an error function measuring the extent to which object S is more similar to object G than to object B; and (e) obtaining an optimal coefficient w by back propagation through the neural network which minimizes the error outputs of the error function.