面向电压降的忆阻神经网络精度优化

Optimization on IR-Drop Induced Accuracy Loss for Memristor-Based Neural Network

由于忆阻器交叉阵列自身的模拟特性可高效实现乘累加运算,因此,它被广泛用于构建神经形态计算系统的硬件加速器.然而,纳米线电阻的存在,会引起忆阻器与纳米线构成的电阻网络出现电压降问题,导致忆阻器阵列的输出信号损失而影响神经网络的精度.分析忆阻器电压降与忆阻器状态、位置,输出电流和输出位置的关系,通过稀疏映射优化电压降,并采用输出补偿进一步提高输出精度.仿真实验的结果表明,该方法可以有效地解决电压降引起的问题,忆阻神经网络在手写数字数据集MNIST的识别率达到95.8%,较优化前提升了33.5%.

The analog properties of memristor cross array(MCA)can efficiently realize multiplication and accumulation(MAC)operation.Hence,MCA is widely used to construct the hardware accelerator of neuromorphic computing system.However,due to the nanowire resistance,the resistive network composed of the memristor and the nanowire suffers from IR-drop,which causes unavoidable loss in the output and hence affects the accuracy of neural network.In this paper,the relationships between the IR-drop of the memristor and its state,position,output current and output position are analyzed.Then,IR-drop is optimized by sparse mapping and output compensation is employed to further improve output accuracy.Experiments show that the optimization strategies proposed can effectively solve the IR-drop induced problem,and the recognition accuracy of the memristor-based neural network on MNIST dataset reaches 95.8%,with 33.5%improvement compared to the pre-optimization.