Parallel Training for Back Propagation in Character Recognition

Artificial Neural Network has made the character recognition work easier and they grow tremendously in improving accuracies and efficiency. However, there are always gaps and weaknesses which is need to prevail due to recognition inaccuracies and less discussed especially in handling large scale of data. The parallelism can be regarded as the practical solution in solving large workload. In achieving an optimal training time and generalization ability, possessing the problem for generating suitable comprehensive classifier will affected positively to the time and maintaining the accuracy in the same time. This paper presents an idea of distributing data the same neural network structure to measure the capability of time reduced for recognition accuracy. The MNIST benchmark dataset is used represent handwritten digits for recognition. To test the validity, the data sets are handled in parallel computers in separate training of back propagation neural network. The results indicate the proposed algorithm improve the speed- up performance for large scale neural networks while maintaining its accuracies