Bayesian Inference Based Robust Computing on Memristor Crossbar
Hosted in Virtual Platform
Emerging Device Technologies
DescriptionMemristor based crossbar is a promising platform for neural network acceleration. To deploy a trained network model on memristor crossbar, memristors need to be programmed to realize the expected weights of the network. However, due to device variation and analog noise, weight deviation from the expected value is inevitable and inference accuracy thus degrades. In this paper, we propose a Bayesian inference based learning framework that incorporates variation models into prior weight distribution and transforms it to a posterior distribution to accommodate perturbations. Simulation results with the proposed framework confirm stable inference accuracy under device variation and noise.