Noise-based local learning using stochastic magnetic tunnel junctions

Abstract

Brain-inspired learning in physical hardware has enormous potential for rapid learning with minimal energy expenditure. One of the characteristics of biological learning systems is their ability to learn in the presence of various noise sources. Inspired by this observation, we introduce a noise-based learning approach for physical systems implementing multilayer neural networks. Simulation results show that our approach allows for effective learning with a performance approaching that of the conventional backpropagation algorithm, which is effective but has a high energy cost. Using a spintronics hardware implementation, we demonstrate experimentally that learning can be achieved in a small network composed of physical stochastic magnetic tunnel junctions. These results provide a path toward efficient learning in general physical systems that embraces rather than mitigates the noise inherent in physical devices.