A Low Energy-Barrier Magnet based Analog Stochastic Neuron

Neuromorphic hardware built using emerging nano-materials technology can provide enormous scalability and energy-efficiency over conventional digital circuit designs for both edge and cloud AI devices. In this work, we propose a compact energy-efficient Analog Stochastic Neuron device built from a low energy-barrier magnetic tunnel junction and a few silicon transistors that can function as a compact drop-in replacement for stochastic sigmoidal neurons in a “software” model of a neural network (NN), allowing building of large scale NN circuits where neurons are first class objects. We discuss in details the physics that allows us to build such a device and using a comprehensive coupled stochastic magneto-dynamics and charge and spin transport, demonstrate multiple applications of this device using illustrative examples.