On Performance and Calibration of Natural Gradient Langevin Dynamics

Abstract

Producing deep neural network (DNN) models with calibrated confidence is essential for applications in many fields, such as medical image analysis, natural language processing, and robotics. Modern neural networks have been reported to be poorly calibrated compared with those from a decade ago. The stochastic gradient Langevin dynamics (SGLD) algorithm offers a tractable approximate Bayesian inference applicable to DNN, providing a principled method for learning the uncertainty. A recent benchmark study showed that SGLD could produce a more robust model to covariate shifts than other competing methods. However, vanilla SGLD is also known to be slow, and preconditioning can improve SGLD efficacy. This paper proposes eigenvalue-corrected Kronecker factorization (EKFAC) preconditioned SGLD (EKS-GLD), in which a novel second-order gradient approximation is employed as a preconditioner for the SGLD algorithm. This approach is expected to bring together the advantages of both second-order optimization and the approximate Bayesian method. Experiments were conducted to compare the performance of EKSGLD with existing preconditioning methods and showed that it could achieve higher predictive accuracy and better calibration on the validation set. EKSGLD improved the best accuracy by 3.06% on CIFAR-10 and 4.15% on MNIST, improved the best negative log-likelihood by 16.2% on CIFAR-10 and 11.4% on MNIST, and improved the best thresholded adaptive calibration error by 4.05% on CIFAR-10.