Double magnetic tunnel junction based ∑Δ∑ hardware neuron

Brain inspired spiking neurons have gained traction in edge machine learning (ML) applications as they encode information in temporal spikes improving energy efficiency [1] . However, replicating certain characteristics of spiking neurons, like their stochastic variations in firing threshold, is difficult in conventional silicon devices. Here we propose a novel double magnetic tunnel junction (MTJ) based spiking neuron which leverages the stochasticity of low barrier magnets (LBM) to emulate the firing threshold variability. The neural encoding schemes used in spiking neurons have recently started using schemes like Delta [2] and Sigma-delta modulation [3] . These facilitate the communication of features between neurons only when they surpass a certain threshold. As such, we use an encoding scheme termed as Sigma-Delta-Sigma (SDS) from our earlier work on temporal computation of information [4] . A unique feature of SDS encoding is its capability to filter out noise-like features from relevant information ensuring the noise-robustness of the network, a large improvement over previous LBM-based neurons.