Evaluation of memory capacity of spin torque oscillator for recurrent neural networks
Sumito Tsunegi, Tomohiro Taniguchi, Shinji Miwa, Kohei Nakajima, Kay Yakushiji, Akio Fukushima, Shinji Yuasa, Hitoshi Kubota · Japanese Journal of Applied Physics · 2018
Abstract We have succeeded in experimentally observing the memory functionality of a vortex-type spin torque oscillator used in recurrent neural networks for the first time. Voltage pulses representing random binary pattern of 2000 (1000) bits were applied to the oscillator for learning (test), and the amplitude of the oscillator was evaluated as the 200 virtual nodes. The memory capacity of the whole system including the oscillator was evaluated to be 1.8 at maximum. The value is larger than the memory capacity of the system without the oscillator, indicating the existence of a finite contribution from the spin torque oscillator.