We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Statistics > Machine Learning

Title: Non-convex Learning via Replica Exchange Stochastic Gradient MCMC

Abstract: Replica exchange Monte Carlo (reMC), also known as parallel tempering, is an important technique for accelerating the convergence of the conventional Markov Chain Monte Carlo (MCMC) algorithms. However, such a method requires the evaluation of the energy function based on the full dataset and is not scalable to big data. The na\"ive implementation of reMC in mini-batch settings introduces large biases, which cannot be directly extended to the stochastic gradient MCMC (SGMCMC), the standard sampling method for simulating from deep neural networks (DNNs). In this paper, we propose an adaptive replica exchange SGMCMC (reSGMCMC) to automatically correct the bias and study the corresponding properties. The analysis implies an acceleration-accuracy trade-off in the numerical discretization of a Markov jump process in a stochastic environment. Empirically, we test the algorithm through extensive experiments on various setups and obtain the state-of-the-art results on CIFAR10, CIFAR100, and SVHN in both supervised learning and semi-supervised learning tasks.
Comments: Accepted by ICML 2020
Subjects: Machine Learning (stat.ML); Machine Learning (cs.LG); Probability (math.PR); Methodology (stat.ME)
Cite as: arXiv:2008.05367 [stat.ML]
  (or arXiv:2008.05367v3 [stat.ML] for this version)

Submission history

From: Wei Deng [view email]
[v1] Wed, 12 Aug 2020 15:02:59 GMT (8119kb,D)
[v2] Thu, 10 Sep 2020 02:28:35 GMT (8119kb,D)
[v3] Mon, 22 Mar 2021 15:55:25 GMT (8147kb,D)

Link back to: arXiv, form interface, contact.