Optimization of Reservoir Computing Utilizing Interfered Spin Waves

Abstract

Listen

Translate article

Reservoir computing is a learning model that enables low-cost and fast learning compared to conventional deep learning. It enables physical implementation by replacing the reservoir part with a physical device that possesses nonlinearity, short-term memory, and high dimensionality. In particular, it was theoretically predicted that spin wave interference can perform highly efficient reservoir computing in micromagnetic simulations,[1] and it was experimentally verified by the present authors.[2] Whereas the computational performance was suggested to vary significantly depending on the interval of input voltage pulses used for spin wave excitation and the applied magnetic field intensity, the details were unclear to date. Therefore, we expanded the measurement range to investigate detailed mechanism of the system and to explore untouched high performance in the study.We use a device fabricated on the surface of a single crystal, as shown in Fig. (a), then we performed one of the most important prediction benchmark tasks, the Nonlinear Autoregressive Moving Average (NARMA) model. Since it has been found that utilizing the nonlinear interference of spin waves through multi-detection demonstrates high performance, in this study, we utilized both Detection A and Detection B. We evaluated the prediction error of NARMA10, which requires short-term memory up to 10 steps, by varying the applied magnetic field and the interval of input voltage pulses under measurement conditions. As shown in Fig. (b), the comparison between the target waveform and the output waveform during training and testing under the conditions of a magnetic field of 182 mT and a pulse interval of 23.6 ns shows a good match, at that time, the error was 0.183 during the testing phase. Next, we examined the detailed condition dependence in NARMA10. As indicated in Fig. (c), high performance is obtained in the region where the pulse interval is between 20 and 30 ns, while the performance was low in the region below 19.3 ns, particularly in the strong magnetic field range from 188 mT to 200 mT. We will discuss the condition dependence with the other NARMA models like NARMA2. Additionally, we will discuss in detail the dependence on input pulse intervals, along with evaluations of short-term memory capacity and nonlinearity. This work was in part supported by Innovative Science and Technology Initiative for Security Grant Number JPJ004596, ATLA, Japan and JST PRESTO (JPMJPR23H4).【Reference】[1] Nakane et al., IEEE Access 6, 4462 (2018).[2] Namiki et al., Adv. Intell. Syst. 5, 2300228 (2023). Figure 1