基于数字神经元仿生器件的类脑计算技术研究
2022-11-27 14:02:10
论文总字数:19279字
摘 要
伴随着信息时代的到来,当今社会的数据成爆发式方式快速增长,各种新兴技术对数据处理的速度和效率有了更高的要求。传统的冯诺依曼计算架构因存算分离导致的问题,无法适应新时代快而准的要求了;而传统的Cmos器件尺寸正在逼近物理极限,摩尔定律难以延续。同时,现代化人工智能针对目标识别等功能的迫切需求,如今的深度学习已难以解决功耗,算力等问题。因此,类脑神经网络计算技术成为解决这些问题的关键,类脑计算通过拟神经网络架构实现局域存算融合,在无人驾驶、语音和图像识别等智能化任务处理应用场景下,有着广泛的应用前景,并且其在功耗、能效、算力方面的优势显著,有望成为大数据及人工智能时代的引领者。
本文介绍了如下内容:(1)数字神经元的基本定义;(2)神经形态计算和深度学习的内涵以及二者的基本模型和架构,以及近年来二者间相互发展的趋势;(3)概述了新一代人工神经元仿生器件的发展,并着重选取了国内外具有代表性的几款芯片,例如国内清华大学的天机芯,英特尔公司的loihi芯片,IBM的truenorth芯片,美国斯坦福大学的neurogrid芯片,浙江大学的达尔文芯片,分别从其基本原理,架构,算法等方面进行简单介绍;(4)最后对基于数字神经元仿生器件的类脑计算技术进行总结和展望。
关键词:神经形态器件;人工智能;类脑计算;神经形态芯片
Research on brain like computing technology based on digital neuron bionic device
ABSTRACT
With the advent of the information age, the data of today's society is growing rapidly in an explosive way, and various emerging technologies have higher requirements for the speed and efficiency of data processing. The traditional von Neumann computing architecture can not meet the requirements of fast and accurate in the new era because of the problems caused by the separation of storage and computation; The size of traditional CMOS devices is approaching the physical limit, and Moore's law is difficult to continue. At the same time, modern artificial intelligence for the urgent needs of target recognition and other functions, today's deep learning has been difficult to solve the problems of power consumption, computing power and so on. Therefore, brain like neural network computing technology becomes the key to solve these problems. Brain like computing realizes local memory computing fusion through quasi neural network architecture. It has a wide range of application prospects in unmanned driving, voice and image recognition and other intelligent task processing application scenarios, and has significant advantages in power consumption, energy efficiency and computing power, It is expected to become a leader in the era of big data and artificial intelligence.
This paper introduces the following contents: (1) the basic definition of digital neuron; (2) the connotation of neural morphological computing and deep learning, as well as their basic model and architecture, and the mutual development trend in recent years;(3) This paper summarizes the development of a new generation of artificial neuron bionic devices, and focuses on several representative chips at home and abroad, such as tianjixin chip of Tsinghua University, Loihi chip of Intel company, truenorth chip of IBM, neurogrid chip of Stanford University, Darwin chip of Zhejiang University, The algorithm and other aspects are briefly introduced;(4) Finally, the brain like computing technology based on digital neuron bionic device is summarized and prospected.
Keywords: neuromorphological devices;artificial intelligence;brain like computing;neuromorphological chips
目 录
摘 要 I
ABSTRACT II
第一章 绪论 1
1.1 背景与意义 1
1.2 研究内容 1
1.3 文章基本框架 1
第二章 数字神经元仿生器件的基础 3
2.1 神经元的组成 3
2.2 神经元传递信号的过程 3
2.3 神经元的基本模型 4
2.3.1 Hodgkin-Huxley模型 4
2.3.2 LIF模型 4
2.3.3 Izhikevich模型 5
2.3.4 SRM模型 5
2.4 数字神经元的定义 5
2.5 数字神经元的架构 5
2.5.1 通用近存计算架构 5
2.5.2 SRAM存算一体 6
2.5.3 DRAM存算一体 7
2.6 本章小结 7
第三章 神经形态计算与深度学习 8
3.1 神经形态计算的内涵 8
3.2 深度学习的内涵 8
3.3 基本模型 8
3.3.1 面向神经科学的模型 8
3.3.2 深度学习模型 8
3.3.3 模型上的差异 9
3.4 发展趋势 9
3.5 混合范式架构 10
3.6 本章小结 10
第四章 以数字神经元为基础的类脑芯片 11
4.1 天机芯 11
4.1.1 天机芯片规格 11
4.1.2 FCore架构 12
4.1.3 天机芯片的优缺点 12
4.2 IBM的truenorth 12
4.2.1 Truenorth芯片规格 12
4.2.2 Truenorth架构 13
4.2.3 Truenorth芯片的优缺点 15
4.3 英特尔的loihi 15
4.3.1 Loihi芯片规格 15
4.3.2 Loihi芯片架构 16
4.3.3 Loihi芯片优缺点 16
4.4 斯坦福大学Neurogrid芯片 17
4.4.1 芯片规格 17
4.4.2 Neurogid芯片优缺点 17
4.5 本章小结 17
第五章 总结与展望 18
致 谢 19
参考文献 20
第一章 绪论
1.1 背景与意义
自从20世纪50年代发明出集成电路以来,人们跟随着摩尔定律不断地提高集成电路的集成度,推动着冯诺依曼架构计算机性能持续提升。冯诺依曼结构计算机具有处理器模块和存储器模块分离的特征,其通过预先存储的不同软件程序指令依次执行,但是这一特点也使得处理计算和存储数据的频繁调换,导致产生了“存储墙”问题,并且难以对实际场景和需求的不同进行实时的自我改变,实现智能化发展。
剩余内容已隐藏,请支付后下载全文,论文总字数:19279字