Decision making module based on stochastic magnetic tunnel junctions

Abstract

In biological neural systems, noise is ubiquitous but does not affect the correct decisions made in the complex cognitive tasks. Decision-making in biological neural system is typically achieved by accumulating input information over a period of time. Inspired by recent developments in neurosciences, we design a decision-making module based on spintronic devices, utilizing superparamagnetic tunnel junctions as artificial neurons. The feasibility of this decision-making module is verified through circuit simulations. Taking a multi-layer perceptron as an example, the module significantly improves the accuracy of the perceptron in the handwritten digit recognition task. Furthermore, the spintronic decision-making module offers advantages over the conventional pooling methods, such as adaptive decision time, high performance and the absence of analog-to-digital conversion. The decision-making module is flexible to be integrated into artificial neural networks and provides a general yet effective solution to enhance performance against device noise.