Lightweight Neural Network Technology for Edge Computing

Abstract

With the increase in computational requirements for deep learning, it has become difficult for these models to be applied to the development of edge devices in the Internet of Things; thus, lightweight networking technology has been one of the main research directions recently. This paper will present an organised list of light neural network technologies and introduce the three main model compression techniques: pruning, quantization and knowledge distillation. According to the actual engineering conditions, this paper will discuss the three main problems preventing the use of the model, analyze in detail the inherent deficiencies of the current technology: Trade-off between model accuracy and compression ratio, difficulty of edge hardware, limitation between theoretical and actual engineering application. Finally, it puts forward some specific suggestions to find practical ways to improve future lightweight networking technology for edge computing. The existing compression workflow must support more CPU architectures, also the collaborative software-hardware co-design research should continue, and simple edge evaluation should be introduced. This paper provides references for lightweight edge research.