Threshold Determination for ARTMAP-FD Familiarity Discrimination

Abstract

The ARTMAP-FD neural network performs both identification (placing test patterns in classes encountered during training) and familiarity discrimination (judging whether a test pattern belongs to any of the classes encountered during training). ARTMAP-FD quantifies the familiarity of a test pattern by computing a measure of the degree to which the pattern's components lie within the ranges of values of training patterns grouped in the same cluster. This familiarity measure is compared to a threshold which can be varied to generate a receiver operating characteristic (ROC) curve. Methods for selecting optimal values for the threshold are evaluated. The performance of validation-set methods is compared with that of methods which track the development of the network's discrimination capability during training. The techniques are applied to databases of simulated radar range profiles.