A NEW CORNER - CLASSIFICATION APPROACH TO NEURAL-NETWORK TRAINING

The corner classification approach to neural network training has the excellent capability of prescriptive learning, where the network weigh ts are prescribed merely by inspection of the training samples. This t echnique is extremely fast compared to other conventional training tec hniques such as backpropagation. However, the versions described hithe rto have been sensitive to the choice of the radius of generalization. We present here a new and improved corner classification technique th at retains the prescriptive learning capability and gives excellent ge neralization performance. This algorithm could be the basis of the rec ently introduced neuroscientific notion of ''working memory.''.