Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: A Hitting Time Analysis of Stochastic Gradient Langevin Dynamics
(Submitted on 18 Feb 2017 (v1), last revised 9 Apr 2018 (this version, v3))
Abstract: We study the Stochastic Gradient Langevin Dynamics (SGLD) algorithm for non-convex optimization. The algorithm performs stochastic gradient descent, where in each step it injects appropriately scaled Gaussian noise to the update. We analyze the algorithm's hitting time to an arbitrary subset of the parameter space. Two results follow from our general theory: First, we prove that for empirical risk minimization, if the empirical risk is point-wise close to the (smooth) population risk, then the algorithm achieves an approximate local minimum of the population risk in polynomial time, escaping suboptimal local minima that only exist in the empirical risk. Second, we show that SGLD improves on one of the best known learnability results for learning linear classifiers under the zero-one loss.
Submission history
From: Yuchen Zhang [view email][v1] Sat, 18 Feb 2017 06:33:55 GMT (192kb,D)
[v2] Tue, 6 Jun 2017 18:59:05 GMT (246kb,D)
[v3] Mon, 9 Apr 2018 07:14:34 GMT (234kb,D)
Link back to: arXiv, form interface, contact.