Lightweight Deep Neural Network Models for Edge Intelligence:A Survey
With the rapid development of the Internet of Things(IoT)and artificial intelligence(AI),the combination of edge computing and AI has given rise to a new research field called edge intelligence.Edge intelligence possess appropriate computing power and can provide real-time,efficient,and intelligent responses.It has significant applications in areas such as smart cities,in-dustrial IoT,smart healthcare,autonomous driving,and smart homes.In order to improve the accuracy of models,traditional deep neural networks often adopt deeper and larger architectures,resulting in significant increases in model parameters,storage re-quirements,and computational complexity.However,due to the limitations of IoT terminal devices in terms of computing power,storage space,and energy resources,deep neural networks are difficult to be directly deployed on these devices.Therefore,light-weight deep neural networks with low memory,low computational resources,high accuracy,and real-time inference capability have become a research hotspot.This paper first reviews the development process of edge intelligence and analyzes the practical requirements for lightweight deep neural networks to adapt to intelligent terminals.Two methods for constructing lightweight deep neural network models:model compression techniques and lightweight architecture design are proposed.Next,it discusses in detail five main model compression techniques:parameter pruning,parameter quantization,low-rank decomposition,knowledge distillation,and mixed compression techniques.It summarizes their respective performance advantages and limitations,and eva-luates their compression effects on commonly used datasets.Then,the paper analyzes in depth the strategies of adjusting the size of the convolution kernel,reducing input channel number,decomposing convolution operations,and adjusting convolution width in lightweight architecture design,and compares several commonly used lightweight network models.Finally,the future research di-rection of lightweight deep neural networks in the field of edge intelligence is prospected.
NETL
NSTL
万方数据
