Enhanced High Dimensionality and the Information Processing Capacity in Interfered Spin Wave‐Based Reservoir Computing, Achieved With Eight Detectors

ABSTRACT Physical reservoir computing (PRC) is a promising neuromorphic computing framework to significantly reduce the computational resources required for machine learning, although computational performance has plenty of room for improvement compared to simulation‐based machine learning models. In this study, we fabricated a spin wave interference‐based PRC device with ten terminals to systematically investigate the effect of the number of detectors, from one to eight detectors, on computational performance. The 10‐step‐ahead prediction task of chaotic time series data generated by the Mackey‐Glass equation was performed, and the eight‐detector device achieved a significantly improved computational performance with a root mean squared error (RMSE) of 1.63 × 10 − 2 . This result represents top‐level performance among currently reported PRCs and demonstrates exceptionally high performance compared to simulation‐based machine learning models. Information processing capacity (IPC) analysis evidenced that the spin wave interference‐based PRC with eight detectors has outstanding linear capacity, which has a strong negative correlation with the RMSE of the chaotic time series prediction. This study confirmed in these systems that multi‐terminalization is an effective technique capable of significantly improving computational performance through the enhancement of high dimensionality.