Stochastic neuron based on dual-free-layer magnetic tunnel junction

Abstract

In this Letter, we present a stochastic neuron implemented based on a dual-free-layer magnetic tunnel junction (MTJ). The device exhibits a stochastic switching behavior by utilizing coupled degrees of freedom and dynamic oscillation induced by spin transfer torque in both free magnetic layers. The stochasticity can be effectively modulated by an external magnetic field and bias current. Furthermore, we exploit this stochastic behavior to emulate a sigmoid activation function, which we employ for handwritten digit recognition on the Mixed National Institute of Standards and Technology database, achieving a recognition accuracy of 95.14%, comparable to that of conventional activation functions. Our work demonstrates a promising approach for constructing artificial stochastic neurons based on dual-free-layer MTJ, offering richer dynamic properties for neuromorphic computing applications.