Revolutionizing real-time data processing with edge computing and reservoir technology

Every day, large amounts of data related to weather, traffic, and social media are processed in real time. In traditional cloud computing, this processing is performed in the cloud, which raises concerns about issues such as information leaks, communication delays, slowdowns, and increased power consumption.

Against this background, “edge computing” presents a promising alternative solution. It is located close to the user and aims to reduce the load and speed up data processing by distributing computations. Specifically, edge AI, which processes AI at the edge, is expected to be applied to things such as self-driving cars and predicting abnormalities in factory machinery.

However, effective edge computing requires efficient and computationally cost-effective technologies. One promising option is reservoir computing, a computational technique designed to process signals recorded over long periods of time. Reservoirs that respond nonlinearly to these signals can be used to transform these signals into complex patterns.

In particular, physical reservoirs that use physical system dynamics are both computationally expensive and efficient. However, the ability to process signals in real time is limited by the natural relaxation time of physical systems. This limits real-time processing and requires tuning for best learning performance.

Recently, Professor Kentaro Kinoshita of the Department of Applied Physics, Faculty of Advanced Engineering, Tokyo University of Science, Yutaro Yamazaki of the Graduate School of Science, and others have developed an optical system. A device with the ability to support physical reservoir computing and enable real-time signal processing over a wide range of timescales within a single device. Their discovery is cutting edge science On November 20, 2023.

Professor Kinoshita said, “The device developed in this research will be able to process time-series signals of various time scales that occur in our daily life environment in real time with a single device.Especially in the edge domain. AI devices to utilize.”

In their research, the two created a special device using Sn-doped In.₂○₃ and Nb-doped SrTiO₃ (denoted as ITO/Nb:STO), responds to both electrical and optical signals. They tested the electrical characteristics of the device and confirmed that it functions as a memristor (a memory device that can change electrical resistance). The research team also investigated the effect of UV light on ITO/Nb:STO by varying the voltage and observing the changes in current. The results suggest that this device can change the relaxation time of the photoinduced current depending on the voltage, making it a potential candidate for a physical reservoir.

Additionally, the team tested its effectiveness as a physical reservoir using ITO/Nb:STO for classification of handwritten digit images in the Modified National Institute of Standards and Technology (MNIST) dataset. To our delight, the device achieved a classification accuracy of up to 90.2%. Additionally, to understand the role of physical reservoirs, the team performed experiments without physical reservoirs, which resulted in a relatively low classification accuracy of 85.1%. These findings demonstrate that the ITO/Nb:STO bonded device improves classification accuracy while keeping computational costs low and proves its value as a physical reservoir.

“Our research group has focused on research and development of materials that can be applied to physical reservoir computing.Therefore, with the aim of realizing a physical reservoir in which the relaxation time of photoinduced current can be arbitrarily controlled, we “We created the device by applying voltage,” says Professor Kinoshita.

In summary, this study presents a new memristor device that can tune the response timescale through voltage changes and exhibits enhanced learning ability, making it promising for applications at the edge as an AI device for edge computing. This could pave the way for a single device that can effectively process signals of different durations found in real-world environments.