Implementation of an oversize neural network on DAP-510

摘要

The authors propose parallel implementation of a multi-layerfeedforward neural network on the AMT DAP-510. The network is trainedusing the Back Propagation algorithm. The proposed implementation is foran oversize network, that is for a network which has more numbers ofneurons than the processors. Also, the implementation has theflexibility of varying the size of the network. The peak performance of2.6 million interconnections per second is achieved for a three-layernetwork with 512 neurons in each layer. They compare their results withthe existing parallel implementations