The parallel processing of neural networks on ring-shaped multiprocessor

Abstract

An efficient parallel simulation algorithm is described for a multilayered neural network with a backpropagation learning procedure on a distributed-memory multiprocessor with ring connection. Each layer is divided into P disjoint blocks and each block is mapped on each processor of a P-processor system. The parallel simulation using this algorithm has been implemented on a system of transputers with ring connection, and the time complexity of both calculation and communication is analyzed. It is shown that the algorithm can provide high efficiency on a multiprocessor system which has a large number of processors connected in a ring.<>