氧化物神经元器件及其神经网络应用  

Oxide Neuron Devices and Their Applications in Artificial Neural Networks

作　　者：李宗晓 胡令祥 王敬蕊[2] 诸葛飞[1,3,4,5] LI Zongxiao;HU Lingxiang;WANG Jingrui;ZHUGE Fei(Ningbo Institute of Materials Technology and Engineering,Chinese Academy of Sciences,Ningbo 315201,China;School of Electronic and Information Engineering,Ningbo University of Technology,Ningbo 315211,China;Center for Excellence in Brain Science and Intelligence Technology,Chinese Academy of Sciences,Shanghai 200031,China;Center of Materials Science and Optoelectronics Engineering,University of Chinese Academy of Sciences,Beijing 100029,China;Institute of Wenzhou,Zhejiang University,Wenzhou 325006,China)

机构地区：[1]中国科学院宁波材料技术与工程研究所,宁波315201 [2]宁波工程学院电子与信息工程学院,宁波315211 [3]中国科学院脑科学与智能技术卓越创新中心,上海200031 [4]中国科学院大学材料与光电研究中心,北京100029 [5]浙江大学温州研究院,温州325006

出　　处：《无机材料学报》2024年第4期345-358,共14页Journal of Inorganic Materials

基　　金：国家自然科学基金(U20A20209);中国科学院战略性先导专项(XDB32050204);中国博士后创新人才支持计划(BX2021326);中国博士后科学基金(2021M703310);浙江省自然科学基金(LQ22F040003);宁波市自然科学基金(2021J139,2023J356);环境友好能源材料国家重点实验室开放基金(20kfhg09)。

摘　　要：目前,人工智能在人类社会发挥着越来越重要的作用,以深度学习为代表的人工智能算法对硬件算力的要求也越来越高。然而随着摩尔定律逼近极限,传统冯·诺依曼计算架构越来越难以满足硬件算力提升的迫切需求。受人脑启发的新型神经形态计算采用数据处理与存储一体架构,有望为开发低能耗、高算力的新型人工智能技术提供重要的硬件基础。人工神经元和人工突触作为神经形态计算系统的核心组成部分,是当前研究的前沿和热点。本文聚焦氧化物人工神经元,从神经元数学模型出发,重点介绍了基于氧化物电子器件的霍奇金–赫胥黎神经元、泄漏–累积–发射神经元和振荡神经元的最新研究进展,系统分析了器件结构、工作机制对神经元功能模拟的影响规律。进一步,根据不同尖峰发射动态行为,阐述了基于氧化物神经元硬件的脉冲神经网络和振荡神经网络的研究进展。最后,讨论了氧化物神经元在器件、阵列、神经网络等层面面临的挑战,并展望了其在神经形态计算等领域的发展前景。Nowadays,artificial intelligence(AI)is playing an increasingly important role in human society.Running AI algorithms represented by deep learning places great demands on computational power of hardware.However,with Moore's law approaching physical limitations,the traditional von Neumann computing architecture cannot meet the urgent demand for promoting hardware computational power.The brain-inspired neuromorphic computing(NC)employing an integrated processing-memory architecture is expected to provide an important hardware basis for developing novel AI technologies with low energy consumption and high computational power.Under this conception,artificial neurons and synapses,as the core components of NC systems,have become a research hotspot.This paper aims to provide a comprehensive review on the development of oxide neuron devices.Firstly,several mathematical models of neurons are described.Then,recent progress of Hodgkin-Huxley neurons,leaky integrate-and-fire neurons and oscillatory neurons based on oxide electronic devices is introduced in detail.The effects of device structures and working mechanisms on neuronal performance are systematically analyzed.Next,the hardware implementation of spiking neural networks and oscillatory neural networks based on oxide artificial neurons is demonstrated.Finally,the challenges of oxide neuron devices,arrays and networks,as well as prospect for their applications are pointed out.

关 键 词：氧化物 神经元器件 类脑计算 神经形态计算 人工神经网络 综述 

分 类 号：O472[理学—半导体物理]