Research map / Probabilistic and scientific modeling / Variational and generative sampling
Pathfinder: Parallel quasi-Newton variational inference
Lu Zhang, Bob Carpenter, Andrew Gelman, Aki Vehtari
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace ADVI (mean-field or dense) with single-path Pathfinder | posterior approximation · posteriordb (20 models) | 1-Wasserstein distance: median distance of mean-field ADVI more than double for 8 of 20 models; dense ADVI 9 of 20 | Improved |
| use Pathfinder instead of Stan phase I warmup | posterior approximation · posteriordb (20 models) | 1-Wasserstein distance: 7 of 20 models with more than double the distance | Improved |
| use Pathfinder instead of Stan phase I and ADVI | computational cost · posteriordb | log density and gradient evaluations: gradient evaluations 34, 48, 54 times more than Pathfinder | Improved |
| use multi-path Pathfinder instead of single-path | Pathfinder · bball_drive_event_0-hmm_drive_0 | 1-Wasserstein distance | Improved |
| reduce L-BFGS iterations to 100 and loosen tolerance to 10^-3 (ablation) | Pathfinder · posteriordb | 1-Wasserstein distance: increased more than a factor of 8 for three models | Hurt |
| increase ELBO Monte Carlo draws from K=5 to K=30 (ablation) | Pathfinder · posteriordb | 1-Wasserstein distance: 3.2% average, 15.1% maximum reduction; 4.9 times log-density evaluations | Improved |
| increase L-BFGS history from J=6 to J=60 (ablation) | Pathfinder · posteriordb | 1-Wasserstein distance: 0.7% average reduction for 19 models; maximum reduction 16.3%, maximum increase 9.8% | Mixed |
| decrease number of single-path runs from I=20 to I=5 (ablation) | multi-path Pathfinder · posteriordb | 1-Wasserstein distance: median increases 5.4% on average, 0.1% minimum, 12.5% maximum | Hurt |
| increase number of single-path runs from I=20 to I=40 (ablation) | multi-path Pathfinder · posteriordb | 1-Wasserstein distance: reduced by 1.1% on average, maximum reduction 5.5%, maximum increase 6.4% | No clear effect |
| use more diffuse initialization uniform(-15,15) instead of uniform(-2,2) (ablation) | multi-path Pathfinder · eight schools centered / 100-d Neal funnel | approximate draw coverage | Improved |
| use ADVI instead of Pathfinder | posterior approximation · eight schools centered | 1-Wasserstein distance and ELBO | Improved |
Closest papers by meaning
- Understanding and Accelerating Particle-Based Variational Inference (ICML 2019 · similarity 0.90)
- Black-Box Variational Inference as a Parametric Approximation to Langevin Dynamics (ICML 2020 · similarity 0.89)
- Particle Semi-Implicit Variational Inference (NeurIPS 2024 · similarity 0.88)
- Doob's Lagrangian: A Sample-Efficient Variational Approach to Transition Path Sampling (NeurIPS 2024 · similarity 0.88)
- Least squares variational inference (NeurIPS 2025 · similarity 0.87)
- Transport meets Variational Inference: Controlled Monte Carlo Diffusions (ICLR 2024 · similarity 0.87)
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.
