Synchronous Unsupervised STDP Learning with Stochastic STT-MRAM Switching*

The use of analog resistance states for storing weights in neuromorphic systems is impeded by fabrication imprecision and device stochasticity that limit the precision of synapse weights. This challenge can be resolved by emulating analog behavior with the stochastic switching of the binary states of spin-transfer torque magnetoresistive random-access memory (STT-MRAM) [1]. STT is a stochastic process that switches the MTJ state with a probability dependent on the pulse voltage and duration. We propose a synchronous system in which input neurons generate input pulses to the MRAM array and output neurons integrate current from the MRAM array. The output neurons perform the leaking, integration, and firing functions. This synchronous circuit design has been demonstrated via behavioral simulation and the inference accuracy can reach 90% on MNIST handwritten digits. These results are comparable to simulations of unsupervised single layer SNNs based on multilevel memristors evaluated with a similar size and methodology [2]. The proposed binary STT-MRAM system with stochastic writing will soon be experimentally proven to provide higher accuracies than can be achieved with memristors and phase-change memory.

[1] A. Vincent et al., IEEE TBioCAS, 2015.

[2] N. Rathi et al., IEEE TCAD, 2018.