ON PARALLELIZING RECURSIVE NEURAL NETWORKS ON COARSE-GRAINED PARALLELCOMPUTERS - A GENERAL ALGORITHM

This paper describes a parallel algorithm that simulates the retrievin g phase of recursive neural networks. The neural network is divided in to sub-networks, each mapped on a coarse-grained processor. Such a coa rse-grained approach, requiring the partitioning of the neural network into units significantly larger than neurons, guarantees that the per formance does not depend on the updating mode. The performance obtaine d on a transputer-based machine is also shown and commented.