A Review of Research on Convolutional Neural Network Model Compression and Hardware Acceleration for Edge Intelligence

Abstract

With the combination of Internet of Things (IoT) and 5G (Fifth Generation Mobile Networks) technologies, a new intelligent computing method, Edge Intelligence, has emerged, which has the characteristics of less delay, fully protected privacy and little power consumption. However, there are a lot of parameters and calculations in convolutional neural networks, but there are many problems such as the amount that edge devices can store, the intensity of calculation, and the lack of available electricity. So far, all the comments are about the model company and hardware acceleration as one thing, and there is a lack of in-depth and broad summary about the software and hardware working together. This paper is a comment on the key technologies of model compression and hardware acceleration in edge intelligence, and it has always focused on the fact that software and hardware work together. In the model compression part, the hardware acceleration part, Graphics Processing Units (GPUs), Field-Programmable Gate Arrays (FPGAs) and Application-Specific Integrated Circuits (ASICs)/Neural Processing Units (NPUs) are summarized from four aspects: network pruning, parameter quantization, knowledge distillation and lightweight networks, and the hardware features and matching methods for model synthesis are summarized. On this basis, this paper introduces many latest research results of collaborative optimization. It also includes hardware-aware compression search, dynamic inference scheduling, cross-platform automatic adaptation and so on. Then it also talks about the important issues of collaborative optimization and the future development direction. There are also detailed and easy-to-understand quantity comparison tables. It summarizes the advantages and disadvantages of various compression algorithm in performance, hardware adaptability and engineering technology. These aims are to help the heterogeneous design and development of edge vision terminals such as autonomous driving, industrial inspection and smart cameras.