Accelerating Convergence of Replica Exchange Stochastic Gradient MCMC via Variance Reduction
Abstract
Replica exchange stochastic gradient Langevin dynamics (reSGLD) has shown promise in accelerating the convergence in non-convex learning; however, an excessively large correction for avoiding biases from noisy energy estimators has limited the potential of the acceleration. To address this issue, we study the variance reduction for noisy energy estimators, which promotes much more effective swaps. Theoretically, we provide a non-asymptotic analysis on the exponential convergence for the underlying continuous-time Markov jump process; moreover, we consider a generalized Girsanov theorem which includes the change of Poisson measure to overcome the crude discretization based on the Gr\"owall's inequality and yields a much tighter error in the 2-Wasserstein ($\mathcal{W}_2$) distance. Numerically, we conduct extensive experiments and obtain state-of-the-art results in optimization and uncertainty estimates for synthetic experiments and image data.
Cite
Text
Deng et al. "Accelerating Convergence of Replica Exchange Stochastic Gradient MCMC via Variance Reduction." International Conference on Learning Representations, 2021.Markdown
[Deng et al. "Accelerating Convergence of Replica Exchange Stochastic Gradient MCMC via Variance Reduction." International Conference on Learning Representations, 2021.](https://mlanthology.org/iclr/2021/deng2021iclr-accelerating/)BibTeX
@inproceedings{deng2021iclr-accelerating,
title = {{Accelerating Convergence of Replica Exchange Stochastic Gradient MCMC via Variance Reduction}},
author = {Deng, Wei and Feng, Qi and Karagiannis, Georgios P. and Lin, Guang and Liang, Faming},
booktitle = {International Conference on Learning Representations},
year = {2021},
url = {https://mlanthology.org/iclr/2021/deng2021iclr-accelerating/}
}