Non-convex learning via Stochastic Gradient Langevin Dynamics: a nonasymptotic analysis

Maxim Raginsky, Alexander Rakhlin, Matus Telgarsky
Proceedings of the 2017 Conference on Learning Theory, PMLR 65:1674-1703, 2017.

Abstract

Stochastic Gradient Langevin Dynamics (SGLD) is a popular variant of Stochastic Gradient Descent, where properly scaled isotropic Gaussian noise is added to an unbiased estimate of the gradient at each iteration. This modest change allows SGLD to escape local minima and suffices to guarantee asymptotic convergence to global minimizers for sufficiently regular non-convex objectives. The present work provides a nonasymptotic analysis in the context of non-convex learning problems, giving finite-time guarantees for SGLD to find approximate minimizers of both empirical and population risks. As in the asymptotic setting, our analysis relates the discrete-time SGLD Markov chain to a continuous-time diffusion process. A new tool that drives the results is the use of weighted transportation cost inequalities to quantify the rate of convergence of SGLD to a stationary distribution in the Euclidean $2$-Wasserstein distance.

Cite this Paper


BibTeX
@InProceedings{pmlr-v65-raginsky17a, title = {Non-convex learning via Stochastic Gradient Langevin Dynamics: a nonasymptotic analysis}, author = {Raginsky, Maxim and Rakhlin, Alexander and Telgarsky, Matus}, booktitle = {Proceedings of the 2017 Conference on Learning Theory}, pages = {1674--1703}, year = {2017}, editor = {Kale, Satyen and Shamir, Ohad}, volume = {65}, series = {Proceedings of Machine Learning Research}, month = {07--10 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v65/raginsky17a/raginsky17a.pdf}, url = {https://proceedings.mlr.press/v65/raginsky17a.html}, abstract = {Stochastic Gradient Langevin Dynamics (SGLD) is a popular variant of Stochastic Gradient Descent, where properly scaled isotropic Gaussian noise is added to an unbiased estimate of the gradient at each iteration. This modest change allows SGLD to escape local minima and suffices to guarantee asymptotic convergence to global minimizers for sufficiently regular non-convex objectives. The present work provides a nonasymptotic analysis in the context of non-convex learning problems, giving finite-time guarantees for SGLD to find approximate minimizers of both empirical and population risks. As in the asymptotic setting, our analysis relates the discrete-time SGLD Markov chain to a continuous-time diffusion process. A new tool that drives the results is the use of weighted transportation cost inequalities to quantify the rate of convergence of SGLD to a stationary distribution in the Euclidean $2$-Wasserstein distance.} }
Endnote
%0 Conference Paper %T Non-convex learning via Stochastic Gradient Langevin Dynamics: a nonasymptotic analysis %A Maxim Raginsky %A Alexander Rakhlin %A Matus Telgarsky %B Proceedings of the 2017 Conference on Learning Theory %C Proceedings of Machine Learning Research %D 2017 %E Satyen Kale %E Ohad Shamir %F pmlr-v65-raginsky17a %I PMLR %P 1674--1703 %U https://proceedings.mlr.press/v65/raginsky17a.html %V 65 %X Stochastic Gradient Langevin Dynamics (SGLD) is a popular variant of Stochastic Gradient Descent, where properly scaled isotropic Gaussian noise is added to an unbiased estimate of the gradient at each iteration. This modest change allows SGLD to escape local minima and suffices to guarantee asymptotic convergence to global minimizers for sufficiently regular non-convex objectives. The present work provides a nonasymptotic analysis in the context of non-convex learning problems, giving finite-time guarantees for SGLD to find approximate minimizers of both empirical and population risks. As in the asymptotic setting, our analysis relates the discrete-time SGLD Markov chain to a continuous-time diffusion process. A new tool that drives the results is the use of weighted transportation cost inequalities to quantify the rate of convergence of SGLD to a stationary distribution in the Euclidean $2$-Wasserstein distance.
APA
Raginsky, M., Rakhlin, A. & Telgarsky, M.. (2017). Non-convex learning via Stochastic Gradient Langevin Dynamics: a nonasymptotic analysis. Proceedings of the 2017 Conference on Learning Theory, in Proceedings of Machine Learning Research 65:1674-1703 Available from https://proceedings.mlr.press/v65/raginsky17a.html.

Related Material