Analog spin-orbit torque devices for artificial neural networks

摘要

Development of dedicated neuromorphic hardware operating with compact size, low power consumption, and high endurance is of urgent importance for realization of large-scale and adaptive artificial neural networks (ANNs) that mimic the human brain. Emerging analogue non-volatile memories have attracted a great deal of attention as they can store the synaptic weights in the ANNs, enabling brain-like low-power operation. Spintronics devices offering non-volatility and virtually unlimited endurance are one of the attractive candidates for the artificial synapse in the ANNs. As a proof-of-concept demonstration, we here show an associative memory operation using spin-orbit torque devices with an antiferromagnet-ferromagnet bilayer system that store analogue synaptic weights in the Hopfield model in a non-volatile fashion: We confirm that device mismatching can be compensated by learning process, demonstrating the capability of analogue spintronics devices as an artificial synapse in ANNs.