Bayesian inference using spintronic technology: A proposal for an MRAM-based stochastic logic gate

Bayesian inference forms the basis for many models of human cognition as well as in the development of autonomous systems that must rapidly make decisions based upon ambiguous or incomplete sensory data. Recent research has shown that a simple Muller C-element can implement Bayes' rule directly using stochastic computation. In this work, the design of a stochastic Muller C-element that uses spin-based magnetic tunnel junction devices is described. These nanodevices allow the compact implementation of the memory buffers, which are required to reshuffle the output bitstream in order to minimize the problem of autocorrelation inherent to the Muller C gate. A simple rotating buffer is shown to be effective in mitigating the impact of autocorrelation when used with a multi-input C element. Improved area efficiency is expected compared to previous designs as the control signals for the memory buffer can be shared across several C-elements. Circuit simulations and numerical analyses are presented to demonstrate the operation of this gate.