Artificial spiking neurons may form an element base for the next generation of neuromorphic computing devices. A promising design of an ultra-fast artificial neuron is based on antiferromagnetic (AFM) spin Hall oscillators driven by a sub-threshold spin current (1). These devices could produce ultra-short voltage spikes in response to a weak external stimulus (1). Fixed synapses can connect these AFM neurons in simple neuromorphic circuits (2). To create more complex neural networks, it is necessary to find learning algorithms that are compatible with AFM neurons. Here, we study the problem of AFM neural network training for pattern recognition tasks. We use reservoir computing, such that only the weights connected to the output neuron are altered during training. A supervised machine learning algorithm based on the temporal position of the spikes, called spike pattern association neuron (SPAN) (3), is used during training. Synaptic weight adjustments are governed by the difference between spikes' desired and actual temporal position. An AFM SPAN, trained to recognize a symbol made from a grid, will produce a spike at a target time when the symbol is supplied as input. SPANs, trained to recognize different symbols, are connected to the same inputs. An output layer is created to ensure that only the spike corresponding to the recognized symbol is sent to the output. This output layer consists of a clock neuron spiking at the target time and weakly fixed synapses (Fig. 1). To create an output spike, a SPAN must produce a spike at the target time along with the clock neuron (Fig. 2). Using the SPAN algorithm, we could develop a neural network with AFM artificial neurons capable of performing the pattern recognition tasks.

References:
(1) R. Khymyn et al. Sci. Rep. 8, 15727 (2018).
(2) H. Bradley and V. Tyberkevych, Q3-01, MMM, 2020.
(3) A. Mohemmed, et al., Neurocomputing, vol 107, pp 3-10, 2013.