発表者名 |
小林 海翔 |
指導教員名 |
求 幸年 教授 |
発表題目(英語) |
Thermally-robust spatiotemporal parallel reservoir computing in frustrated magnets |
要旨(英語) |
Physical reservoir computing is a cutting-edge framework for neuromorphic machine learning that exploits the nonlinear phenomena in a physical system to achieve low-power, high-speed, and versatile hardware implementation for real-time machine learning [1]. Magnetic materials are one of the most promising platforms for physical reservoir computing, utilizing the nonlinearity and high dimensionality of complicated spin dynamics [2]. Meanwhile, two major obstacles remain for device applications of the spintronic physical reservoirs. One is robustness against thermal fluctuations, which is crucial for retaining short-term memory, the essence of neuromorphic computing. The other is that parallel processing techniques, a key aspect for high integration, remain largely unexplored.
In this study, we demonstrate, for a simple model of frustrated magnets, both robustness to thermal fluctuations and feasibility of frequency division multiplexing [3]. We find that when the input is provided through an AC magnetic field, the information is retained in the spin dynamics at around the same frequency as the input AC field, even though the overall memories are quickly disrupted by thermal noise. This observation motivates the use of a frequency filter to remove thermal noise from irrelevant frequencies, leading to the preservation of the short-term memory even at finite temperatures. Moreover, we achieve parallel processing with frequency division multiplexing by utilizing a superposed magnetic field that carries multiple information at different frequencies. Our scheme can be coupled with parallelization in spatial domain at the level of a single spin, yielding a vast number of spatiotemporal computational units. Furthermore, we show that the nonlinearity arising from the exchange interaction allows information processing between different frequency threads, including linearly-inseparable logic gate tasks such as XOR and XNOR, without the need for dedicated communication channels. Our framework can be implemented on a variety of magnetic materials, paving a way for the device realization of physical reservoir computing [4].
[1] H. Jaeger and H. Haas, Science 304, 78-80 (2004).
[2] J. Torrejon et al., Nature 547, 428-431 (2017).
[3] K. Kobayashi and Y. Motome, arXiv:2302.14496 (2023).
[4] 求幸年、小林海翔、特願2023-25883 「情報処理システム、情報処理方法およびプログラム」 |
発表言語 |
日本語 |