Efficient and Flexible Edge Computing

Physical reservoir computing can be used to perform high-speed processing for artificial intelligence with low power consumption.

Japanese researchers design a physical tunable reservoir device based on dielectric relaxation at an electrode-ionic liquid interface.

In the near future, more and more AI processing will need to take place at the edge – close to the user and where the data is collected rather than at a remote computer server. This will require high speed data processing with low power consumption. Physical reservoir computing is an attractive platform for this purpose, and a new breakthrough from Japanese scientists has just made this much more flexible and practical.

Physical Reservoir Computing (PRC), which relies on the transient response of physical systems, is an attractive machine learning framework that can perform high-speed processing of low-power time-series signals. However, PRC systems have poor tunability, which limits the signals they can process. Now, Japanese researchers present ionic liquids as an easily tunable physical reservoir that can be optimized to process signals over a wide range of timescales simply by changing their viscosity.

Artificial intelligence (AI) is rapidly becoming pervasive in modern society and will be implemented more widely in the years to come. In applications involving sensors and IoT devices, the standard is often advanced artificial intelligence, a technology in which computation and analytics are performed close to the user (where data is collected ) and not far away on a centralized server. This is because edge AI has low power requirements as well as high-speed data processing capabilities, characteristics that are particularly desirable for processing real-time time-series data.

Time scale of signals commonly produced in living environments

Time scale of signals commonly produced in living environments. The response time of the liquid ion PRC system developed by the team can be tuned to be optimized for processing such real-world signals. Credit: Kentaro Kinoshita of TUS

In this regard, Physical Reservoir Computing (PRC), which relies on the transient dynamics of physical systems, can greatly simplify the edge AI computing paradigm. Indeed, the PRC can be used to store and process analog signals in these edges that the AI ​​can work and analyze efficiently. However, the dynamics of solid PRC systems are characterized by specific time scales which are not easily tunable and are generally too fast for most physical signals. This shift in time scales and their poor controllability make PRCs largely unsuitable for real-time signal processing in living environments.

To solve this problem, a research team from Japan involving Professor Kentaro Kinoshita and Sang-Gyu Koh, a doctoral student, from Tokyo University of Science, and principal investigators Dr. Hiroyuki Akinaga, Dr. Hisashi Shima and Dr. Yasuhisa Naitoh of the National Institute of Advanced Industrial Science and Technology, proposed, in a new study published in the journal Scientific reports, the use of liquid PRC systems instead. “Replacing conventional solid reservoirs with liquid reservoirs should lead to AI devices capable of directly learning on time scales from signals generated by the environment, such as voice and vibrations, in real time,” explains Professor Kinoshita. “Ionic liquids are stable molten salts that consist entirely of free electrical charges. The dielectric relaxation of ionic liquid, or the way its charges rearrange in response to an electrical signal, could be used as a reservoir and holds great promise for advanced AI physics computation.

Tank calculation based on ionic liquid

The response of the ionic liquid PRC system can be tuned to be optimized for processing a wide range of signals by altering its viscosity by adjusting the length of the cationic side chain. Credit: Kentaro Kinoshita of TUS

In their study, the team designed a PRC system with an ionic liquid (IL) of an organic salt, 1-alkyl-3-methylimidazolium bis(trifluoromethanesulfonyl)imide ([Rmim+] [TFSI] R = ethyl (e), butyl (b), hexyl (h) and octyl (o)), whose cationic part (the positively charged ion) can be easily varied with the length of a chosen alkyl chain. They fabricated gold slit electrodes and filled the gaps with IL. “We found that the time scale of the reservoir, although complex in nature, can be directly controlled by the viscosity of IL, which depends on the length of the cationic alkyl chain. Changing the alkyl group in organic salts is easy to do and presents us with a controllable and conceivable system for a range of signal lifetimes, enabling a wide range of computational applications in the future,” Prof. Kinoshita explains. By adjusting the length of the alkyl chain between 2 and 8 units, the researchers obtained characteristic response times between 1 and 20 µs, with longer alkyl side chains leading to longer response times and better performance. adjustable AI learning devices.

The tunability of the system was demonstrated using an AI image identification task. The AI ​​received a handwritten image as input, which was represented by rectangular pulse voltages 1 µs wide. By increasing the length of the side chain, the team brought the transient dynamics closer to that of the target signal, with the discrimination rate improving for longer chain lengths. This is because, compared to [emim+] [TFSI]in which the current relaxed to its value in about 1 µs, the IL with a longer side chain and, in turn, a longer relaxation time better preserved the time series data history, improving identification precision. When the longest side chain of 8 units was used, the discrimination rate reached a maximum value of 90.2%.

Input signal conversion via ionic liquid-based PRC system

Conversion of the input signal via the PRC system based on ionic liquid. The tank output as a current response (top and middle) to an input voltage pulse signal (bottom) is shown. If the current decay (dielectric relaxation) is too fast/slow, it reaches its saturation value before the next signal input and no history of the previous signal is kept (middle image). Whereas, if the current response attenuates with a relaxation time that correctly matches the time scales of the input pulse, the history of the previous input signal is retained (top image). Credit: Kentaro Kinoshita of TUS

These results are encouraging as they clearly show that the proposed PRC system based on dielectric relaxation at an electrode-ionic liquid interface can be suitably tuned according to the input signals by simply changing the viscosity of the IL. This could pave the way for advanced artificial intelligence devices capable of accurately learning the various signals produced in the living environment in real time.

IT has never been so flexible!

Reference: “Reservoir computing with dielectric relaxation at an electrode–ionic liquid interface” by Sang-Gyu Koh, Hisashi Shima, Yasuhisa Naitoh, Hiroyuki Akinaga and Kentaro Kinoshita, April 28, 2022, Scientific reports.
DOI: 10.1038/s41598-022-10152-9

Kinoshita Kentaro is a professor in the Department of Applied Physics, Tokyo University of Science, Japan. His area of ​​interest is device physics, with a focus on memory devices, AI devices, and functional materials. He has published 105 articles with over 1600 citations to his credit and holds a patent in his name.

This study was partially funded by JSPS KAKENHI Grant Number JP20J12046.

Tokyo University of Science (TUS) is a well-known and respected university, and the largest private research university specializing in science in Japan, with four campuses in central Tokyo and its suburbs and in Hokkaido. Founded in 1881, the university has continuously contributed to the scientific development of Japan by instilling a love of science in researchers, technicians and educators.

Sherry J. Basler