A parallel design and implementation for backpropagation neural network using MIMD architecture

The backpropagation (BP) neural network is characterized by a slow rate of convergence learning speed especially in case of large architecture and large training data sets. So, this paper illustrates a parallel BP neural network depends upon MIMD system architecture using transputers. A parallel design version of BP neural network offers an attractive alternative for improving the learning speed. The paper introduces a mechanism to decrease the communication overhead between different related processor elements to reduce the communication time of parallel processors. The developed parallel design of BP neural network is experimented with the problem of printed Arabic character recognition. The experimental results revealed that the speed up approaches the maximum value of 15.2 when a large number of training patterns is used and the size of the network is large too.