AI Paper Atlas

Research map / Optimization / Optimizers

On the geometry of Stein variational gradient descent

Andrew Duncan, Nikolas Nüsken, Lukasz Szpruch

JMLR 2023 · Optimization › Optimizers · approach: Stein variational gradient descent

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
decrease kernel smoothness parameter p from 2 to 1 in the exponential-power kernel (ablation)SVGD with k_{p,sigma}, 500 particles · 1D four-component Gaussian mixtureparticle approximation accuracy (histogram, Wasserstein-1)Improved
decrease p further below 1 (to 1/2 and smaller) (ablation)SVGD with k_{p,sigma}, 500 particles · 1D four-component Gaussian mixtureWasserstein-1 error and gradient evaluationsHurt
use p=1 instead of p=2 or p=1/2 at fixed p, measured per gradient evaluation (ablation)SVGD with k_{p,sigma} · 1D four-component Gaussian mixtureWasserstein-1 error vs gradient evaluationsImproved
anneal p from p0=2 to p1=1/2 over time (log-linear schedule) instead of fixed p (ablation)SVGD with k_{p,sigma} · 1D four-component Gaussian mixtureWasserstein-1 errorMixed
use p=0.5 instead of p=2 or p=1 (ablation)SVGD with k_{p,sigma}, 500 particles · 2D six-component Gaussian mixtureWasserstein-1 errorImproved
use non-translation-invariant weighted Matern kernel with adapted tailsSVGD · numerical studiescomputational cost vs accuracyNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy