南京大学学报（自然科学版）2025，Vol.61Issue(5)：867-878,12.DOI:10.13232/j.cnki.jnju.2025.05.015
基于1T1R忆阻器交叉阵列与CMOS激活函数的全模拟神经网络
Fully Analog Neural Network based on 1T1R memristor crossbar array and CMOS activation functions
摘要
Abstract
Neuromorphic circuits based on memristor crossbar arrays offer a highly promising technological route for energy-efficient implementation of neural network computation.However,existing schemes often require extensive analogue-to-digital conversion processes,leading to an efficiency bottleneck.This paper proposes a fully analog neural network architecture based on a 1T1R(1 Transistor 1 Resistor)memristor crossbar array and CMOS(Complementary Metal-Oxide-Semiconductor)activation functions,as well as a customized training method.By utilizing the 1T1R memristor crossbar array to achieve analog computation of matrix multiplication,the continuous tunable conductance of the memristors is leveraged to map neural network weights,and the parallel multiply-accumulate operations are performed in accordance with Ohm's Law and Kirchhoff's Law.Additionally,diverse CMOS analog circuits,such as pseudo-ReLU(Rectified Linear Unit),Sigmoid,and Tanh,are designed to implement nonlinear activation functions.Through the customized training method that optimizes analog hardware characteristics,an accuracy rate of 98%is achieved on the MNIST(Modified National Institute of Standards and Technology)handwritten digit recognition task.Our results demonstrate the feasibility of the proposed architecture despite the programming noise of memristors,providing a new approach to improving the computational efficiency.关键词
全模拟神经网络/忆阻器/类脑电路/CMOS激活函数/1T1R交叉阵列Key words
fully analog neural network/memristor/Brain-inspired circuit/CMOS activation function/1T1R crossbar array分类
信息技术与安全科学引用本文复制引用
赵航,杨董行健,王聪,梁世军,缪峰..基于1T1R忆阻器交叉阵列与CMOS激活函数的全模拟神经网络[J].南京大学学报（自然科学版）,2025,61(5):867-878,12.基金项目
国家自然科学基金(62034004,62305155,62122036),江苏省自然科学基金(BK20232004,BK20233001),中国博士后基金(2023M731582,BX20230153) （62034004,62305155,62122036）
