Parallelizing convolutional neural network for the handwriting recognition problems with different architectures
Junhao Zhou, Weibin Chen, Guishen Peng, Hong Yu Xiao, Hao Henry Wang, Zhigang Chen · 2017
As the convolutional neural network (CNN) algorithm is proved to be uncomplicated in the image preconditioning and relatively simple train the original image, it has become popular in image classification. Apart from the field of image classification, CNN has been widely used in many scientific area, especially in the field of pattern classification. In this paper, we use CNN for handwritten numeral recognition. The basic idea of our method is to use the multi-process to process the training samples in parallel, to exchange the training results and to get the final weight parameters. Compared with the conventional algorithm, the training time is greatly reduced, and the result can be obtained more quickly. Besides, the accuracy of the algorithm is proved to be almost the same as that of the conventional algorithm with sufficient training testing samples. This significantly improves the efficiency of CNN in the hand written numeral recognition. Finally, we also implemented our proposed method with parallel acceleration optimization based on Many Integrated Core Architecture (MIC) architecture of Intel and GPU architecture of Nvidia.