Bayesian neural network enhancing reliability against conductance drift for memristor neural networks

作     者：Yue ZHOU Xiaofang HU Lidan WANG Shukai DUAN Yue ZHOU;Xiaofang HU;Lidan WANG;Shukai DUAN

作者机构：College of Electronics and Information EngineeringSouthwest University College of Artificial IntelligenceSouthwest University 

出 版 物：《Science China(Information Sciences)》 (中国科学:信息科学(英文版))

年 卷 期：2021年第64卷第6期

页      面：100-111页

核心收录：

学科分类：080903[工学-微电子学与固体电子学] 0810[工学-信息与通信工程] 0808[工学-电气工程] 0809[工学-电子科学与技术（可授工学、理学学位）] 08[工学] 081104[工学-模式识别与智能系统] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported by National Key R&D Program of China (Grant No. 2018YFB1306600) National Natural Science Foundation of China (Grant Nos. U20A20227, 62076207, 62076208, 61976246) Fundamental Science and Advanced Technology Research Foundation of Chongqing,China (Grant No. cstc2017jcyjBX0050) Natural Science Foundation of Chongqing (Grant No. cstc2020jcyj-msxmX0385) 

主　　题：conductance drift neuromorphic computing Bayesian neural network memristor crossbar array network reliability 

摘      要：The hardware implementation of neural networks based on memristor crossbar array provides a promising paradigm for neuromorphic computing. However, the existence of memristor conductance drift harms the reliability of the deployed neural network, which seriously hinders the practical application of memristor-based neuromorphic computing. In this paper, the impact of different types of conductance drift on the weight realized by memristors is investigated and analyzed. Then, utilizing the weight uncertainty introduced by conductance drift, we propose a weight optimization method based on the Bayesian neural network, which can greatly improve the network performance. Furthermore, an ensemble approach is proposed to enhance network reliability without increasing training cost or crossbar array resources. Finally, the effectiveness of the proposed scheme is verified through a series of experiments. In addition, the proposed scheme can be easily integrated into the implementation of neuromorphic computing, which can provide a better guarantee for its large-scale application.