A parallel computing and mathematical method optimization of CNN network convolution

Training convolutional neural network (CNN) is a compute-intensive task in parallel tolerance has become a complete training is very important. There are two obstacles in the distributed memory computing environment to develop a scalable parallel CNN. It depends on a small volume across the two model parameters shown adjacent to the height data. The other presents the maximum overlapping parallel computing inter-process communication, large amounts of data over a communication channel to go. They will be transferred to the calculation. Replication by using threads on each compute node to initiate communication available after gradient is achieved. Reverse spread output data is generated in each stage of the model layer, and data communication may be performed in parallel with the calculation of other layers. To impact the replication study's efficiency and scalability, evaluated the model structure and optimization of various mathematical methods. When using the image VGGnet model training dataset, use 256 and 512, respectively, small batch size to achieve speedup computing nodes.