Presentation Information
[16a-M_123-5]A Novel Node-Extraction Scheme Enabling High-Performance, Node-Efficient Spin-Wave Reservoir Computing with Hardware Demonstration
〇(P)Jiaxuan Chen1, Ryo Iguchi1, Sota Hikasa1,2, Takashi Tsuchiya1,2 (1.NIMS, 2.Tokyo Univ. of Sci.)
Keywords:
spin waves,physical reservoir computing
Spin-wave reservoir computing exploits the intrinsic nonlinearity and history-dependent dynamics of spin waves for non-von Neumann information processing. Due to the limited number of physical detectors, high-dimensional reservoirs are typically realized via virtual-node extraction, which inevitably increases hardware complexity and imposes operational constraints (such as complicated clocking and digital integrated circuits). In this study, we propose a new node-extraction scheme that not only enhances the effective utilization of information carried in spin waves but also enables simple and scalable hardware implementation suitable for edge devices. We experimentally implement the proposed scheme with analog hardware components. We verify its performance on benchmark tasks including parity check (PC) and nonlinear autoregressive moving average (NARMA). Using fewer than sixty reservoir nodes, we achieve a total capacity of 3.2 for PC, which is the highest value reported to date among experimentally demonstrated physical reservoirs, and a normalized mean square error of 0.0076 for NARMA-2, which is also comparable to those of recently advanced physical reservoirs employing hundreds of nodes. These results pave the way toward the near-term realization of spin-wave reservoir computing as a promising hardware platform for compact, high-performance, and node-efficient neuromorphic computing systems.