Implementation of an oversize neural network on DAP-510

The authors propose parallel implementation of a multi-layer feedforward neural network on the AMT DAP-510. The network is trained using the Back Propagation algorithm. The proposed implementation is for an oversize network, that is for a network which has more numbers of neurons than the processors. Also, the implementation has the flexibility of varying the size of the network. The peak performance of 2.6 million interconnections per second is achieved for a three-layer network with 512 neurons in each layer. They compare their results with the existing parallel implementations.<>