Spike Pattern Recognition Using Antiferromagnetic Artificial Neurons

It has been shown that antiferromagnetic (AFM) spin Hall oscillators driven by a sub-threshold spin current produce ultra-short (∼5 ps) spike in response to an external stimulus and can be used as ultra-fast artificial neurons [1]. These AFM neurons can be connected by passive synapses to create neuromorphic circuits [2]. We show that a neural network based on AFM neurons can use a supervised machine learning algorithm, Spike patten association neuron (SPAN) [3]. This algorithm is based on the Widrow-Hoff learning rule and makes use of the temporal encoding of the neuron’s spikes. SPAN is an example of reservoir computing, as only the weights connected to the output layer are altered. An AFM neural network is trained to recognize letters made from a 9 celled grid of neurons. We use a modified SPAN approach with an inhibitory circuit, made possible by the effective inertia of the AFM neuron [2], as an output layer with constant weights. The inhibitor suppresses any excess unwanted data from the input. This modified SPAN allows for all weights to remain positive, which makes it easier to be implemented in hardware.

 

 

 

[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.