A Summary of convolution Neural Network Compression and Acceleration Technology

摘要

Although convolution neural network has achieved remarkable results in different application scenarios, there are a large number of parameters and computation in its structure, which limit its development in mobile and embedded devices. How to reduce parameters, compress model and optimize structure to improve network performance without losing accuracy has become a hot issue of convolution neural network. This paper summarizes and summarizes the convolution neural network structure optimization technology from five aspects: granularity pruning, weight quantization sharing, knowledge distillation, tensor decomposition and fine network design, and analyzes the technical core of it. Their advantages and disadvantages, applicable scenarios and optimization results are analyzed and summarized respectively, and the future research direction is prospected.