Analysis of perceptron training algorithms and applications to hand-written character recognition

Issues regarding the convergence of training algorithms for perceptron networks are addressed. The algorithms are the perceptron convergence procedure, the back propagation algorithm, and a recently developed modification to the back propagation algorithm, referred to as the selective update back propagation algorithm. It is shown that networks trained with the back propagation algorithm can only be implemented as a read-only memory while those trained with the selective update back propagation algorithm can be used as read-and-write memory. It is further shown that updates of back propagation will never cease while those of selective update back propagation will. To elucidate the theoretical results, the use of these two algorithms are then employed with perceptron networks for application to character recognition is discussed