Online Training of Spiking Neural Networks Using Domain Wall Magnetic Tunnel Junction Synapses

Abstract

Progress has been made on emulating the neuron and the synapse using magnetic devices [1], [2] especially with three-terminal magnetic tunnel junctions (3T-MTJs) [3], [4]. Increasingly more accurate biological mimics are being developed with the 3T-MTJ like the Leaky-Integrate-Fire (LIF) [5] neuron displaying lateral inhibition, and the 3T-MTJ synapse model showing spike timing dependent plasticity (STDP) [6], [7]. However, more work must be done to achieve real-time machine learning. Using an earlier developed synapse circuit [8] and a 3T-MTJ spice model [9], we show the transient behavior of a 2 by 2 crossbar array (cartoon in Fig 1). There is STDP for different delay conditions between the pre- and post-synaptic neural spikes. The shorter the time difference between the onset of the pre- and post-synaptic neuron spike, the higher the current through the ferromagnet (Fig 2). Higher current leads to higher rate of change of the domain wall (DW) position per neural event and higher DW displacement. At t=0, the initial pair of presynaptic and postsynaptic pulses helps to randomly give the DWs a head start. DW for synapse S2 (red plot) rises faster because it has the least delay of 1 ns between its pre- and post-synaptic pulse despite its pulses lagging that of S1 whose DW had a head start. Scaling this up, the 3T MTJ synapse is robust in more complex systems such as a machine learning clustering task.