Fund Project:Project supported by the Jiangsu Specially-Appointed Professor and the Science Research Funds for Nanjing University of Posts and Telecommunications, China (Grant Nos. SZDG20180007, NY217116, KFJJ20170101, NY219014, NY218110), the China Postdoctoral Science Foundation (Grant No. 2018M642290), the National Natural Science Foundation of China (Grant No. 61804079), the University Natural Science Foundation of Jiangsu Province, China (Grant No.18KJD510005), the Innovative Doctor Program, China (Grant Nos. CZ1060619001), and the Graduate Research and Innovation Projects of Jiangsu Province, China (Grant Nos. SJCX19_0256, SJKY19_0811, SJKY19_0806)
Received Date:04 July 2019
Accepted Date:27 August 2019
Available Online:26 November 2019
Published Online:05 December 2019
Abstract: Inspired by the working mechanism of human brain, the artificial neural network attracts great interest for its capability of parallel processing, which is favored by big data task. However, the electronic synapse based on CMOS neural network needs at least ten transistors to realize one biological synaptic function. So, CMOS-based neural network exhibits obvious weakness in speed, power consumption, circuit area and resource utilization and so on, compared with biological synapses. Therefore, how to build neuromorphic circuits and realize biological functions by constructing electronic synapses with low power consumption and high integration density have become the key points for human to realize brain-like computing system. Memristors, as the fourth basic component, is a two-terminal nonlinear device possessing nonlinear conductance that can be tuned continuously. For that special characteristic, it is very similar to biological synapse whose connection strength can be adjusted continuously. In this article, first of all, we study the electrical characteristic of the Cu/MXene/SiO2/W memristor. When applying a positive DC sweeping voltage to the Cu electrode, the Cu electrode is oxidized, generating Cu2+. The generated Cu2+ in function layer tends tomove towards the bottom electrode under the action of electric field. Near the bottom electrode the Cu2+ moving from top electrode are reduced, generating a conductive Cu atom. With Cu atoms accumulating and extending from bottom electrode to top electrode, the memristor is gradually converted from the initial high resistance state (HRS) into the low resistance state (LRS). Secondly, combining with HP model of memristor, we utilize Verilog A language to simulate memristor in the experiment we conducted. Subsequently, we successfully construct the artificial synaptic unit and design the weight differential circuit with self-feedback branch. In the above circuit, we successfully implementa classical " Pavlov's dog” experiment. By applying the sinusoidal signal and pulse signal to the synaptic unit for testing and training it, respectively, the circuit realizes the convention between the conditions that unconditioned stimulus producing unconditioned response to conditioned stimulus producing conditions response. This work takes memristor as a center, through modelling the electrical characteristic of Cu/MXene/ SiO2/W device, we construct a neuromorphic circuit with weight differential branch andself-feedback branch, successfully simulate the classical learning behavior of biological synapses, and realizes the whole process of biologically conditioned reflex, which is illustrated in detail in the experiment on " Pavlov′s dog”. The results will provide effective guidance forconstructing a large scale and high density neuromorphic circuitbased on memristor, thus promoting the realization of brain-like computation in the future. Keywords:memristor/ artificial synapse/ conditioned reflex/ neuromorphiccircuits
全文HTML
--> --> -->
3.测试与仿真结果经典条件反射又称巴普洛夫反射, 其中最著名的实验为巴普洛夫的狗的唾液条件反射. 实验第一阶段, 狗看到食物本能的发生流涎反应, 属于无条件反射行为, 此时食物为非条件刺激(unconditional stimulus, US), 流涎的行为为非条件反应(unconditional reaction, UR); 实验第二阶段, 当没有经过训练, 单独出现铃声时, 狗并不会流涎, 此时铃声为中性刺激(neural stimulus, NS); 实验第三阶段, 在发出铃声的同时提供食物训练, 狗出现流涎反应; 实验第四阶段, 单独发出铃声, 狗经过第三阶段的训练学习后出现流涎反应, 此时流涎属于条件反应(conditional reaction, CR), 铃声由NS转变为条件刺激(conditional stimulus, CS). 为了构造基于忆阻器的神经形态电路, 实现经典条件反射. 首先我们用惠普模型对实验中忆阻器伏安特性进行了拟合, 使用硬件描述语言Verilog A针对忆阻器电学特性对本文中新型二维材料忆阻器进行建模[32,33]. 图2(a)所示为惠普研究团队提出的忆阻器线性漂移模型, 该模型虽能较好地复现忆阻器的连续导态特性, 但在纳米尺度下, 小电压产生的大电场将会进一步在界面附近产生空位的非线性漂移[34], 无法再使用线性漂移模型. 因此, 我们通过增加以下形式的分段线性窗函数$f(x)$解决该边界问题[35]: 图 2 (a)惠普研究小组提出的忆阻器模型; (b)忆阻器模型仿真数据与实验测试数据拟合 Figure2. (a) Memristor model reported by HP group; (b) the fitting of experimental data and the simulation data.
即发生“流涎”事件. 到此实现条件反射全部过程. 图4(a)为实现条件反射时, 输入输出信号的波形变化. 波形包含学习前、学习过程、学习后三个阶段. 学习前: “肉”事件信号输入时, 对应上述分析过程的${U_{{\rm{o1}}}} > 0$, 输出信号与输入信号同相, 判断为“流涎”; “铃声”事件信号输入时, 对应上述分析过程的${U_{{\rm{o}}2}} < 0$, 输出与输入信号反相, 判断为未发生“流涎”, 此时“铃声”为中性刺激. 学习过程: “铃声”、“肉”事件信号同时输入, 输出信号为两支路单独工作时输出信号的叠加, 对应上述分析过程的${U_{{\rm{o}}3}} > 0$. 此时满足条件“铃声”、“肉”事件同时发生, 触发训练模块输出训练信号, 由于电压脉冲能够对忆阻器Ma产生连续调节作用[37-39], 该电路中以方波脉冲作为训练信号作用于铃铛支路, 如该过程中Vout所示. 当满足条件${R_{{{\rm{M}}_{\rm{a}}}}} < {R_{\rm{a}}}$时, 对应上述分析过程的${U_{{\rm{o4}}}} > 0$, 满足当输入只有“铃声”信号时, 输出波形呈现出由负相${U_{{\rm{o}}2}} < 0$到正相${U_{{\rm{o4}}}} > 0$的变化过程, 该变化说明学习过程结束, “铃声”转变为条件刺激. 图 4 (a)仿真过程施加的信号以及各自对应的结果图; (b)对电路中忆阻器模型进行直流电压扫描, 流经忆阻器的电流随仿真时间的变化; (c)训练过程中, 流经忆阻器Ma的电流随时间的变化 Figure4. (a) The signals inputted into the circuit and corresponding output waveform, respectively; (b) the change of current flowing through the memristor against timeduring DC voltage sweeping; (c) the change of current through the memristor Ma during training process.