Reservoir Computing on Spin-Torque Oscillator Array
Taro Kanao, Hirofumi Suto, Koichi Mizushima, Hayato Goto, Tetsufumi Tanamoto, Tazumi Nagasawa · Physical Review Applied · 2019
Reservoir computing (a framework for machine learning) implemented in physical systems has attracted much attention for real-time computing in $e.g.$ time-series modeling and prediction, for which a single spin-torque oscillator (STO) has been proposed. The performance of a single STO will be limited, though. To solve this issue, the authors study reservoir computing on an $a\phantom{\rule{0}{0ex}}r\phantom{\rule{0}{0ex}}r\phantom{\rule{0}{0ex}}a\phantom{\rule{0}{0ex}}y$ of STOs, showing numerically that the system's performance can improve with more STOs, and can become remarkably better than for a standard neural-network model. Interestingly, performance is best near a boundary between synchronized and disordered states, suggesting an enhancement at the edge of chaos.