Stochastic Inference and Learning Enabled by Magnetic Tunnel Junctions

Neuromorphic computational paradigms that exploit the stochastic switching behavior of devices in the presence of thermal noise is bringing about a wave of change in the way we perceive brain-inspired computing. In this article, we present proposals of spintronics enabled neuromorphic computing systems that perform probabilistic inference and online learning. Such stochastic neuromimetic hardware has the potential of enabling a new generation of state-compressed, low-power computing platforms, which can be significantly more efficient and scalable than their deterministic counterparts.