Experimental Comparisons of semi-supervised and supervised ART classifiers

We present a series of experimental results that reveal the merits of semi-supervised learning, when applied to two different types of ARTMAP architectures (Fuzzy ARTMAP and Ellipsoidal ARTMAP). The concept of semi-

supervised learning (SSL) itself was first introduced in the Simplified Boosted ARTMAP architecture by Verzi, et al. 2002, and was extended to Boosted Ellipsoid ARTMAP by Anagnostopoulos, et al, 2002. In the context of SSL, the ARTMAP classifiers are allowed to commit instantaneous misclassification errors by inhibiting their lateral reset activity under certain conditions. This allows for ART classifiers that exhibit an allowable error on the training set. More specifically, these intentional errors depend on the relative frequency, with which patterns of various class labels access ART exemplars (categories), and on the value of a network misclassification tolerance parameter taking values between 0 and 1. A value of an error tolerance of 0 corresponds to fully supervised learning, and a value of an error tolerance of 1 leads to fully unsupervised learning. Our experimental results clearly reflect the advantage of using tolerance values higher than 0. In particular, the experimental results demonstrate that higher degrees of performance generalization can be achieved by means of utilizing SSL, or equivalently by allowing the ART network to commit some error on the training set.