An Efficient and High-Performance Memristor-Based Reservoir Calculation System

Reservoir computing (RC) is an approach to building computer systems inspired by current knowledge of the human brain. Neuromorphic computer architectures based on this approach are composed of dynamic physical nodes, which, combined, can process spatio-temporal signals.

Researchers at Tsinghua University in China recently created a new RC system based on memristors, electrical components that regulate the flow of electric current through a circuit, while recording the amount of charge that passed through it before. This RC system, presented in an article published in Natural electronicsachieved remarkable results, both in terms of performance and efficiency.

“The basic architecture of our RC memristor system comes from our previous work published in Nature Communicationwhere we validated the feasibility of building an analog reservoir layer with dynamic memristors,” Jianshi Tang, one of the researchers who conducted the study, told TechXplore. “In this new work, we further build the analog readout layer with non-volatile memristors. and integrate it with the parallel tank layer based on a dynamic memristor array to implement an all-analog RC system.”

The RC system created by Tang and his colleagues is based on 24 dynamic memristors (DM), which are connected to a physical tank. Its readout layer, on the other hand, is composed of 2048×4 non-volatile memristors (NVM).

“Each DM in the DM-RC system is a physical system with computing power (called a DM node), which can generate rich reservoir states through a time-division multiplexing process,” Tang explained. “These tank states are then fed directly into the NVM network for multiply-accumulate (MAC) operations in the analog domain, yielding the final output.”

Tang and his colleagues evaluated the performance of their dynamic memristor-based RC system by using it to run a deep learning model on two spatiotemporal signal processing tasks. They found that it achieved remarkably high classification accuracies of 96.6% and 97.9% on the arrhythmia detection and dynamic gesture recognition tasks, respectively.

“Compared to the digital RC system, our full analog RC system has equivalent performance in terms of accuracy but saves more than 99.9% of power consumption (22.2 μW vs. 29.4 mW)” , Tang said. “A unique feature of our work is that, to build a complete all-analog RC system, we used two distinct types of memristors: DMs as parallel tanks and NVM arrays as the readout layer, without the aid of digital components, such as those used in the previously reported hardware RC systems.”

The unique system architecture designed by this team of researchers significantly reduces the complexity of RC approaches, while significantly reducing power consumption. In the future, this could thus allow for simpler and larger-scale RC hardware implementations.

“Optimized non-volatile memristors with excellent analog switching characteristics have been integrated to provide end-to-end analog signal transmission and processing throughout the RC system,” Tang said. “Additionally, based on the noise model extracted from our memristor arrays, a noise-sensitive linear regression method was used to train the output weight and effectively mitigate the accuracy loss (less than 2%) caused by the non-ideal characteristics of memristors.”

Tang and his colleagues were the first to demonstrate real-time all-analog signal processing using an RC hardware system. This demonstration finally allowed them to reliably assess the overall power consumption of their system.

“By correlating the experimental data with model simulations, the working mechanism of the DM-RC system, we were also able to learn more about the relationship between the electrical characteristics of physical nodes and system performance,” Tang said. “Specifically, we unveiled two key features (i.e., sill and window) that were extracted from dynamic memristor node characteristics that had a significant impact on reservoir quality.”

After identifying two characteristics that affected the performance of their RC system, Tang and his colleagues were able to define ranges of these two characteristics that led to optimal RC performance. Combined, these ranges and their other findings could serve as a guide for future design and optimization of RC systems. This could help unlock their potential for edge computing, as well as other applications requiring low power consumption and affordable hardware costs.

“In the future, the entire DM-RC system could be miniaturized and monolithically integrated on-chip to further reduce its power consumption and computational latency,” Tang added. “In addition, a deeper and more sophisticated RC system can be built using the DM-RC system as the base unit, which would further improve system performance due to richer tank states and greater memory capacity.”

© 2022 Science X Network