Optimizing Reservoir Computing Based on an Alternating Input-Driven Spin-Torque Oscillator

Neurons in the brain show nonlinear oscillatorlike behaviors, inspiring the development of neuromorphic computing with spin-torque oscillators for their low power consumption and high integration density. The time-multiplexing technique enables utilizing only one device to achieve neuromorphic computing, further reducing power consumption and simplifying device fabrication. However, optimizing such systems remains a challenge and needs further development. A controversial hypothesis claims that dynamic systems possess the optimal computational capacity to address the information at the dynamical phase transition. Here, we apply this hypothesis to improve a reservoir computing system based on an alternating input-driven spin-torque oscillator. The input-driven property of reservoir computing allows us to exploit the dynamics of a spin-torque oscillator by manipulating the input stream. By tuning the ac amplitude in inputs, the system can exhibit diversified dynamic regimes indicated by the largest Lyapunov exponent and echo-state property spectra. Our findings demonstrate that the proper configuration of ac inputs for reservoir computing based on a single spin-torque oscillator can adjust its information-processing capacity profile and enhance its computational performance for specific tasks.