基于自旋电子器件随机性的神经网络认知系统

Nowadays, the development of integrated circuit suffers from many problems such as high power consumption and low reliability, which has been the bottleneck in computing power. Therefore, new information devices and computing methods have stimulated a growing interest. Since spin-transfer torque magnetic RAM (STT-MRAM) based on magnetic tunnel junction (MTJ) devices feature non-volatility, high write and read speed and infinite endurance, STT-MRAM is a critical component to solve the consumption bottleneck. Artificial neural network can work in ultra-low power consumption and high fault-tolerant computing, which is one of the critical components of future smart-chips. This project aims at using magnetic RAM devices in an original probabilistic regime, to demonstrate low-power cognitive-type applications. This project will bring a radically novel point of view that may transform this emerging field. The programming of many nanodevices, and in particular magnetic RAM (MRAM), has an intrinsic random character. Low-energy programming pulses may program the device, but only with a finite probability. We propose to exploit this random aspect of short programming pulses to develop new ultra low power computing paradigms. Our research will focus on Spin Torque Transfer MRAMs. In these devices, the stochastic behavior is indeed controllable. We will aim bioinspired applications, where nanodevices are used similarly to biological synapses, and also other learning systems where synapses learn in ways inspired by machine learning. The final goal of this technology is to develop ultralow power embedded systems capable of extreme adaptation thanks to learning, and capable of processing natural data.The results of this project can have a strong social and economic impact. It is expected that a driver for future electronic devices will be ambient and cognitive intelligent devices that should simplify people's everyday life. They will require electronic systems, which can compute efficiently with the real-life data from sensors with minimum power consumption. Our novel computing paradigm can be a key tool to achieve such systems..

当前集成电路面临高功耗、低可靠性等瓶颈，计算能力难以继续提高，新型信息器件及新型计算方法成为研究热点。由于自旋电子器件具有非易失存储、无限次高速读写、易于与导体工艺集成等优点，被广泛认为是解决功耗瓶颈的关键技术。人工神经网络可以进行超低功耗及高容错性计算，是未来智能芯片的主要计算构架之一。因此本课题基于新型自旋电子器件--自旋转移动量矩非易失磁性随机存储器（STT-MRAM）构建人工神经网路的突触，创造性地利用其在短编程脉冲下固有的、可控的随机特性，实现低功耗认知应用。我们将对磁性隧道结（MTJ）的随机特性进行研究和建模；结合机器学习与仿生学领域的已有成果设计和演示利用该随机特性的突触学习规则；模拟对现实生活中的应用，设计并制造一个小型演示器。本课题将以一种新的方法实现人工神经网络的低功耗认知应用，对人工智能研发和认知应用提供一种十分有益的探索。

受启于人脑的工作机理，神经网络认知系统由于具有大规模并行的三维网络拓扑结构、计算与存储一体化、高能效及固有的容错性和鲁棒性等特性，有望打破传统冯诺依曼架构中的“存储墙”和“功耗墙”问题。本课题面向低功耗、高集成度神经网络认知系统需求，以磁隧道结为基元，从器件建模、电路设计和系统仿真三个层面开展了跨层联合研究，取得的主要研究成果如下：. 1、构建了一个基于自旋转移矩效应的垂直磁各向异性隧道结（STT-PMA-MTJ）的电气简化模型和可用于系统仿真的行为模型，为后续基于自旋电子器件的神经网络认知系统设计及性能分析奠定了基础。. 2、设计了一种基于多个STT-PMA-MTJ并联的复合自旋电子神经突触器件、一种简化随机仿生学习规则及一种自旋随机脉冲神经元电路。基于设计的自旋神经突触器件、自旋神经元电路和学习规则，搭建了一个全自旋脉冲神经网络。系统仿真结果表明设计的全自旋脉冲神经网络可实现高精度神经形态计算，且对器件偏差具有很强的鲁棒性。本项研究的意义在于提出了一种基于二值纳米存储器件来实现神经形态计算的新方法，从而扩充了用于实现神经突触器件和神经元器件的材料系统。. 3、设计了一种基于多个STT-PMA-MTJ垂直堆叠的新型复合自旋电子器件，并从器件物理、器件制备、电路仿真三个角度验证了该器件可模拟单个人工神经突触。此外，基于新型复合自旋电子器件还设计了一种可实现多阈值传递函数的人工神经元电路。仿真结果表明基于设计的新型复合自旋电子器件搭建的全自旋人工神经网络可高精度神经形态计算，且对器件偏差具有很强的鲁棒性。仿真结果分析表明：与人工神经元输出信号精度相比，突触权值精度对全自旋人工神经网络的学习精度的影响更大。. 4、构建了一个用于系统仿真的模拟器，对自旋神经网络认知系统开发提供了一个借鉴的测试平台。. 本课题的研究成果对未来基于自旋电子器件的神经网络认知系统设计有一定的指导作用。