Bayesian Reasoning Enabled by Spin-Orbit Torque Magnetic Tunnel Junctions

TL;DR

The paper addresses the demand for efficient probabilistic reasoning hardware by leveraging spin-orbit torque MTJs (SOT-MTJs) to implement Bayesian network reasoning. It combines a forward-propagating probabilistic neuron network with a simple point-by-point training algorithm that parameterizes CPTs or PDFs without storing large historical data, and demonstrates SOT-MTJs as high-quality probabilistic samplers. Key contributions include a detailed device characterization showing controllable stochastic switching, a 4-node network mapping that links PDFs to network weights, and a real-time training protocol validated on a medical-diagnostic example with demonstrated adaptability to changing data distributions. The work suggests a promising, low-storage hardware pathway for probabilistic neural computation and broadens the applications of spintronic devices in AI tasks.

Abstract

Bayesian networks play an increasingly important role in data mining, inference, and reasoning with the rapid development of artificial intelligence. In this paper, we present proof-of-concept experiments demonstrating the use of spin-orbit torque magnetic tunnel junctions (SOT-MTJs) in Bayesian network reasoning. Not only can the target probability distribution function (PDF) of a Bayesian network be precisely formulated by a conditional probability table as usual but also quantitatively parameterized by a probabilistic forward propagating neuron network. Moreover, the parameters of the network can also approach the optimum through a simple point-by point training algorithm, by leveraging which we do not need to memorize all historical data nor statistically summarize conditional probabilities behind them, significantly improving storage efficiency and economizing data pretreatment. Furthermore, we developed a simple medical diagnostic system using the SOT-MTJ as a random number generator and sampler, showcasing the application of SOT-MTJ-based Bayesian reasoning. This SOT-MTJ-based Bayesian reasoning shows great promise in the field of artificial probabilistic neural network, broadening the scope of spintronic device applications and providing an efficient and low-storage solution for complex reasoning tasks.

Figures (4)

• Figure 1: Characterization of Y-Type SOT-MTJs. (a) Structure of Y-Type SOT-MTJ. (b) SEM image of an MTJ. (c) R-H loop of the Y-type SOT-MTJ obtained with an in-plane field along the easy axis. Field-free switching of the free layer induced by a 50 ns voltage pulse. (d) Relationship between probability and driven voltage. (e - h) Results obtained from continuous measurement under driven voltages of 0.9 V (e), 1.05 V (f), 1.2 V (g).
• Figure 2: The network structure for Bayesian network reasoning. (a) The Bayesian network and conditional probability table (CPT) for generating samples or medical cases. (b) The network used to calculate the probability density function (PDF) corresponding to $v_{i}$.
• Figure 3: The reasoning process of the Bayesian network utilizing the SOT-MTJ.
• Figure 4: (a) Histogram of the state probability of our single-point training algorithm compared to the Bayesian theory. (b) The K-L divergence between our single-point training algorithm and the Bayesian theory decreases with the increase in training cycles. (c) Histogram of the state probability of our single-point training algorithm compared to the Bayesian theory and the accumulated PDF after $10^{5}$ new records. (d) The K-L divergence of our single-point training algorithm (red) or the accumulated PDF (black) with the Bayesian theory from the $10^{5}$ new records. The experimental sampling results (e) and statistical results of generating 1 (f) in the cases of $ABC$ = 100, 101, 110, and 111.