Reservoir Computing (RC) belongs to a new neuromorphic computing paradigm that directly exploits the dynamics and nonlinearities of a physical system for efficient spatiotemporal signal processing. The advantages of low training cost and simple hardware implementation of RC make it a research hotspot in the field of neuromorphic computing. The RC system can be equivalent to a recurrent neural network, but it will have lower hardware overhead and power consumption, because there are no real recursive connections in RC system. This feature of the RC system also eliminates the problem of error accumulation caused by recursive connections, which gives it the advantage of fully analog computing. Most of the existing RC systems are based on fully digital or digital-analog hybrid hardware architecture. These RC systems all require additional analog-to-digital converters and registers to realize data conversion and buffering, which will bring a lot of power consumption and computing latency. In a fully analog RC system, the analog signal can be directly transmitted and processed in the entire system without any data conversion and buffering. However, to realize such a fully analog RC system with extremely low power consumption and hardware overhead, there are still two major problems to be solved: 1. it is necessary to build a state-rich physical reservoir and find the key parameters that determine the performance of the RC system; 2. it is necessary to find ways to reduce the impact of noise during fully analog transmission and processing.
The associate professor Zhong Yanan in FUNSOM together with his co-supervisor has designed a fully analog RC system (DM-RC system) based on two types of memristors, during his post-doctoral research at Tsinghua University. The RC system utilizes 24 dynamic memristors (DMs) to build a physical reservoir, and uses four 2k non-volatile memristor (NVM) arrays to form the readout layer. Each dynamic memristor in the DM-RC system is a physical system with computing power (called a DM node), which can generate rich reservoir states under a special time-multiplexing mechanism. It is found that some key features of DM nodes, such as threshold and window, have a great impact on the performance of DM-RC system. By adjusting the key features of DM nodes, the DM-RC system can achieve optimal performance. In addition, in order to suppress the influence of non-ideal factors such as noise on system performance, the researchers proposed a noise-aware linear regression algorithm, through which the output layer weights with better noise robustness can be trained. On this basis, the researchers applied the DM-RC system to arrhythmia detection and dynamic gesture recognition tasks, achieving high accuracies of 96.6% and 97.9%, respectively. Compared with the digital RC system, the DM-RC system exhibits very close performance, but saves more than 99.9% of power consumption. This work demonstrates the application potential of the fully analog RC system in low-power scenarios such as edge computing and IoT.Relevant results were published in the journal Nature Electronics. Associate Prof. Yanan Zhong from FUNSOM is one of the co-first authors of the paper.
Link to article：https://www.nature.com/articles/s41928-022-00838-3
Title：A memristor-based analogue reservoir computing system for real-time and power-efficient signal processing
Authors：Yanan Zhong#, Jianshi Tang#*, Xinyi Li#, Xiangpeng Liang, Zhengwu Liu, Yijun Li, Yue Xi, Peng Yao, Zhenqi Hao, Bin Gao, He Qian, Huaqiang Wu*
Introduction to Associate Professor Zhong Yanan：http://funsom.suda.edu.cn/23/f5/c4747a467957/page.htm
Project funding: This work was supported in part by China key research and development program (2021ZD0201205), National Natural Science Foundation of China (91964104, 61974081, 62025111, 92064001, 62104126), and the XPLORER Prize.
Editor: Guo Jia