A Procedure to Select the Vigilance Threshold for the ART2 for Supervised and Unsupervised Training

Abstract

The Adaptive Resonance Theory ART2 [1] is used as a non supervised tool to generate clusters. The clusters generated by an ART2 Neural Network (ART2 NN), depend on a vigilance threshold (ρ). If ρ is near to zero, then a lot of clusters will be generated; if ρ is greater then more clusters will be generated. To get a good performance, this ρ has to be suitable selected for each problem. Until now, no technique had been proposed to automatically select a proper ρ for a specific problem. In this paper we present a first way to automatically obtain the value of ρ, we also illustrate how it can be used in supervised and unsupervised learning. The goal to select a suitable threshold is to reach a better performance at the moment of classification. To improve classification, we also propose to use a set of feature vectors instead of only one to describe the objects. We present some results in the case of character recognition.