Current browse context:
stat.ML
Change to browse by:
References & Citations
Statistics > Machine Learning
Title: A Non-Asymptotic Analysis for Stein Variational Gradient Descent
(Submitted on 17 Jun 2020 (v1), last revised 3 Jan 2021 (this version, v4))
Abstract: We study the Stein Variational Gradient Descent (SVGD) algorithm, which optimises a set of particles to approximate a target probability distribution $\pi\propto e^{-V}$ on $\mathbb{R}^d$. In the population limit, SVGD performs gradient descent in the space of probability distributions on the KL divergence with respect to $\pi$, where the gradient is smoothed through a kernel integral operator. In this paper, we provide a novel finite time analysis for the SVGD algorithm. We provide a descent lemma establishing that the algorithm decreases the objective at each iteration, and rates of convergence for the average Stein Fisher divergence (also referred to as Kernel Stein Discrepancy). We also provide a convergence result of the finite particle system corresponding to the practical implementation of SVGD to its population version.
Submission history
From: Anna Korba [view email][v1] Wed, 17 Jun 2020 12:01:33 GMT (78kb,D)
[v2] Mon, 26 Oct 2020 10:12:49 GMT (74kb,D)
[v3] Tue, 27 Oct 2020 07:59:44 GMT (74kb,D)
[v4] Sun, 3 Jan 2021 11:36:09 GMT (72kb,D)
Link back to: arXiv, form interface, contact.