Deep Neural Network Acceleration Method Based on Sparsity.

With the development of deep learning, deep learning has become more and more widely used in artificial intelligence. At this stage, the deep neural network (DNN) based on high-performance GPU and CPU devices has achieved remarkable results in the fields of object detection and recognition. The DNNs have also been applied to social media, image processing and video processing. With the improvement of neural networks, the depth and complexity of various neural networks are also increasing. On the basis of the sparsity of DNN weights, our method analyzes the influence of the weights on the feature map and obtains the relations between convolution layers. The sparsity of the network channel is deduced from the L1 norm and the L2 norm. And the weights of the DNN are pruned according to sparsity. In the vgg-16 experiment, we can accelerate the neural network by 2.7 times without affecting the accuracy of the neural network. Compared to the unstructured pruning, structured pruning based on the sparsity can effectively improve the speed of the forward and backward process, which has a certain significance for the application of DNNs.