Artificial neural components based on magnetic tunnel junctions have emerged as a research hotspot in the field of neuromorphic computing. In this work, we demonstrate dual type artificial neurons constructed using an in-plane magnetic tunnel junction driven by spin torque. By adjusting the direction of the applied direct current, the device can exhibit stochastic switching and nonlinear rectification behaviors, respectively. The former is driven by direct current spin torque, while the latter is driven by both direct current and microwave spin torque. Specifically, the stochastic switching probability and nonlinear rectification voltage as functions of direct current can simulate the sigmoid and ReLU activation functions in neural networks, respectively. Furthermore, the convolutional neural network constructed using these two simulation functions is used for image classification and recognition, with an accuracy rate of 97%. Our work has opened up possibilities for the advancement of artificial neural morphological computing architecture.
Zhang et al. (2025) studied this question.