Differentiable Distribution Model of Stochastic Volatile Memristor-Based Neuron

Abstract

Volatile memristors have received extensive attention in neuromorphic computing. It is of significance to build compact models of these devices for neuromorphic hardware design. In this work, we propose a compact model of metallic filamentary volatile memristor reproducing ideal and stochastic switching behavior in various operation schemes. We further derive the differentiable distribution model of stochastic switching, enabling co-optimization with neural network to obtain optimal operating conditions. Experiments are conducted to validate the compact model with demonstrated neuromorphic application. The compact model bridges memristive characteristics and algorithm optimization for neuromorphic computing.