In-materio reservoir computing based on magnetic skyrmions

Karin Everschor-Sitte, University of Duisburg-Essen, Germany

Learning and pattern recognition necessarily require memory of past events, a property that conventional CMOS hardware must laboriously and artificially emulate. Dynamical systems realized by, for example, various physical matter naturally provide the memory, complexity, and non-linearity required for a wealth of different unconventional computational approaches, including reservoir computing [1]. In contrast to neural networks, physical reservoir computing is a computational paradigm that requires fewer computational resources while considerably reducing computing time. Using simulations, we show that magnetic structures provide a competitive low-power physical implementation of the key ingredient of reservoir computing – the reservoir [2-4]. We developed efficient task-agnostic metrics benchmarking the reservoir’s key features – non-linearity, complexity, and fading memory [5]. We report the highest performance for in-materio reservoir computers for the benchmark test of classifying spoken digits using a multi-input channel magnetic skyrmion reservoir architecture [6]. Our results are on par with state-of-the-art non-in-materio reservoir systems. Experimental realizations of magnetic skyrmion reservoirs further highlight the potential for energy-efficient high-performance skyrmion-based RC [7,8].

[1] Finocchio, et al., arXiv:2301.06727
[2] Prychynenko, et al., Phys. Rev. Appl. 9, 014034 (2018)
[3] Bourianoff, et al., AIP Adv. 8, 055602 (2018)
[4] Pinna, et al., Phys. Rev. Appl. 14, 054020 (2020)
[5] Love, et al., arXiv:2108.01512
[6] Msiska, et al., Adv. Intell. Syst. 5, 2200388 (2023)
[7] Yokouchi, et al., Sci. Adv.8,eabq5652(2022)
[8] Raab et al., Nat. Commun. 13, 6982 (2022)

back to MCIC 2023