Accurate Large-sample Uncertainty Quantification using Stochastic Gradient Markov Chain Monte Carlo

arXiv:2606. 00293v1 Announce Type: new Abstract: Tuning algorithms such as stochastic gradient descent (SGD) and stochastic gradient Langevin dynamics (SGLD) for approximate sampling and uncertainty quantification remains challenging, particularly in the practically relevant settings when the batch size is large or the model is misspecified.