在物联网(IoT,Internet of things)技术迅速发展的背景下,IoT设备受到计算能力、存储空间、通信带宽以及电池寿命的限制,在运行复杂的人工智能(AI,artificial intelligence)算法中,特别是深度学习模型中面临着挑战.模型剪枝技术通过减少神经网络中的冗余参数,在不损伤AI模型性能的前提下可以有效地降低计算和存储需求.该技术适合用于优化部署在物联网设备上的AI模型.首先,回顾了当前流行的结构化剪枝和非结构化剪枝两种典型的模型剪枝技术,两种剪枝技术分别适用于不同的应用场景.之后,详细分析了这些方法在IoT环境下的多样化应用.最后,结合最新研究成果,详细探讨了当前模型剪枝的局限性,并对物联网中模型剪枝方法未来的发展方向进行了展望.
Model pruning techniques in the Internet of things:state of the art,methods and perspectives
In the context of the rapid development of Internet of things(IoT)technology,IoT devices faced challenges in running complex artificial intelligence(AI)algorithms,especially deep learning models,due to the limitations of comput-ing power,storage space,communication bandwidth,and battery life.Model pruning technology could effectively reduce computation and storage requirements by reducing redundant parameters in neural networks without impairing the perfor-mance of AI models.This technique was extremely suitable for optimising AI models deployed on IoT devices.Firstly,two typical model pruning techniques-structured pruning and unstructured pruning,which were currently popular and suit-able for different application scenarios,were reviewed.Secondly,the diverse applications of these methods in IoT envi-ronments were analysed in detail.Finally,the limitations of the current model pruning were discussed in detail in the light of the latest research results,and the future development direction of model pruning methods in IoT was outlooked.
国家科技期刊平台
NSTL
万方数据
