A parallel processing strategy for dynamic learning rate adaptation for feedforward networks

Summary form only given, as follows. The authors discuss heuristics for dynamic learning rate (LR) selection and adaptation for feedforward networks during training. These heuristics are based on competition among a set of parallel processes, each running a version of the backpropagation (BP) training algorithm. Initial LR selection involves use of an embedded BP network which is trained to predict the mean error, given data in the first few hundred epochs of training. These heuristics reduce the training time by up to a factor of seven using training data from an ECG (electrocardiography) application. Simulations were run on a 30 processor Sequent Balance 21000 parallel computer