Multi-Input Programmable Artificial “Neuron” Based on Antiferromagnetic Spintronic Nanostructures

There are many types of artificial neurons capable of mimicking the behavior of biological neurons in a brain, including the recently proposed “neurons” based on antiferromagnetic (AFM) spin Hall oscillators. Such AFM-based “neurons” are capable of generating “Dirac-delta-like” pulses (spikes) of $\sim 1$ ps duration under the action of AC and DC signals. Real-time programming of these devices is still lacking, which is essential for the development of practical AFM-based neural networks. In this paper, we consider an approach to implement an AFM-based programmable “neuron” (“P-neuron”) as a simple two-layer neural network, where each conventional AFM-based “neuron” in the first input layer has only one input. The state of each first-layer “neurons” can be changed independently by applying a particular DC bias current that affects the “neuron’s” working point. All outputs of the first-layer “neurons” are connected to a single multi-input “neuron”, which forms the second layer of “ P-neuron”. We found that by applying different bias currents to different inputs of the “P-neuron” one can change the “sensitivity” of the “P-neuron” to AC signals coming from these signal inputs, i.e., one can implement a programmable AFM-based artificial “neuron”. We believe that the proposed programmable “neurons” and their programming algorithm can be used to develop and optimize ultrafast neural networks based on antiferromagnetic spintronic oscillators.