Stochastic Synapses Made of Magnetic Domain Walls

The stochastic artificial neural network is widely studied nowadays, since it may address the critical issue of uncertainty quantification demanded in many application scenarios. At the hardware level, it calls for devices that can emulate stochastic weight-modification behaviors of synapses with high effi-ciency. In this work, we propose a stochastic spintronic synapse based on spin-torque-induced stochastic magnetic domain-wall (DW) dynamics, including stochastic DW generation by spin-transfer torque and deterministic DW motion induced by spin-orbit torque. The former plays the role of updating the synaptic weights stochastically, while the latter gives rise to multistates, namely, the analog property of the synapse. The proposed DW-based stochastic synapse requires few devices to implement sufficient resistance states. The neural-network-level simulation demonstrates that a spiking neural network based on these stochastic synapses is capable of classifying breast cancer data with a high accuracy of 95.7%.

Automated summary

The stochastic artificial neural network has been widely studied due to its ability to address the critical issue of uncertainty quantification in many applications. This study proposes a stochastic spintronic synapse based on spin-torque-induced stochastic magnetic domain-wall dynamics, which can efficiently emulate the stochastic weight-modification behaviors of synapses and exhibit analog properties. Experimental results demonstrate that a spiking neural network based on these stochastic synapses can classify breast cancer data with a high accuracy of 95.7%.