Fault-tolerance and learning performance of the back-propagation algorithm using massively parallel implementation

Abstract

Mapping the backpropagation (BP) algorithm onto an SIMD (single-instruction-stream, multiple-data-stream) machine, such as the Massively Parallel Processor, is considered. It is shown that the size of the connectionist network underlying BP can be scaled up to large sizes, resulting in improved performance. Specifically, both fault tolerance and learning speed can be enhanced.<>