An adaptive neural network for supervised learning

This paper proposes the architecture of a hybrid Neo-ART/EBP (Adaptive Resonance Theory/Error-Back-Propagation) neural network and describes the results that may be achieved for a specific image vector quantization. Stacking together a simplified input ART layer and an output EBP network allows us to limit the global number of hidden nodes/interconnections and to speed up the convergence time during the training phase. Moreover, in the pattern space, hyperspherical selective attention regions are investigated and the influence of their increasing/decreasing size is discussed.