VARIATIONS ON ADAPTIVE RESONANCE.

In the process of implementing adaptive resonance circuits (ARCs) for a particular application, the authors have previously considered several circuit modifications and alternative processing conditions. They report here on some of these variations. First they examine an adaptive thresholding technique that prevents inadvertent encoding of recognition nodes which can occur when novel patterns are presented. Next, they consider the behavior of an ARC when patterns are iteratively presented for relatively short periods of time. Finally, they discuss the case of continuous ARC operation in which 'neural' activity is not reinitialized with each pattern presentation. The adaptive thresholding technique provides a novel clustering algorithm applicable to both binary and multilevel data.