Citation: Technical Report CAS/CNS TR-99-007, Boston, MA: Boston University.
Abstract: ART and ARTMAP neural networks for adaptive recognition and prediction have been applied to a variety of problems, including automatic mapping from remote sensing satellite measurements, parts design retrieval at the Boeing Company, medical database prediction, and robot vision. This paper features a self-contained introduction to ART and ARTMAP dynamics. An application of these networks to image processing is illustrated by means of a remote sensing example. The basic ART and ARTMAP networks feature winner-take-all (WTA) competitive coding, which groups inputs into discrete recognition categories. WTA coding in these networks enables fast learning, which allows the network to encode important rare cases but which may lead to inefficient category proliferation with noisy training inputs. This problem is partially solved by ART-EMAP, which use WTA coding for learning but distributed category representations for test-set prediction. Recently developed ART models (dART and dARTMAP) retian stable coding, recognition, and prediction, but allow arbitrarily distributed category representation during learning as well as performance.