EFFICIENT NEURAL-NETWORK TRAINING ON A CRAY Y-MP

An efficient implementation of a quasi-Newton algorithm for training f eed-forward neural network on a Gray Y-MP is presented. The most time- consuming step of a neural network training using the quasi-Newton alg orithm is the computation of the error function and its gradient. Para llelization embedded in these computations can be exploited through ve ctorization in a Gray Y-MP supercomputer. We show how they can be carr ied out such that the overall performance of the neural network traini ng process can be enhanced substantially.