Research map / Probabilistic and scientific modeling / Variational and generative sampling
Posterior sampling strategies based on discretized stochastic differential equations for machine learning applications
Frederik Heber, Žofia Trst’anová, Benedict Leimkuhler
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 |
|---|---|---|---|
| add the ensemble quasi-Newton preconditioner with 8 or 16 walkers instead of a single BAOAB walker | BAOAB sampler · Gaussian model sampling | integrated autocorrelation time of slow directions: significantly improves the Integrated Autocorrelation Time for the slow directions | Improved |
| add the ensemble quasi-Newton preconditioner with 8 or 16 walkers | BAOAB sampler · Gaussian model sampling (fast directions) | integrated autocorrelation time of fast directions | No clear effect |
| use covariance blending eta of 10,000 (EQN) instead of eta=0 (standard Langevin) (ablation) | BAOAB sampler with 8 walkers · MNIST two-class (7 vs 9) | slowest integrated autocorrelation time: improvement of up to a factor of 4 | Improved |
| vary covariance blending constant eta up to very high values (ablation) | EQN sampler · MNIST two-class (7 vs 9) | integrated autocorrelation time | No clear effect |
| increase friction gamma and step size, large potential prefactor a=4 versus a=0.01 (ablation) | BAOAB sampler · Harmonic potential kinetic energy error | error vs step size: slopes -0.44 to 0.08 for a=0.01 versus 0.63 to 1.94 for a=4 | Mixed |
| use BAOAB instead of SGLD, GLA1 and GLA2 (ablation) | Langevin sampler choice · Two-cluster classification virial error | accuracy of position marginal: BAOAB reaches the CLT limit while GLA1 slope 0.86, GLA2 2.01 | Improved |
| use second-order GLA2 instead of first-order GLA1 (ablation) | Geometric Langevin Algorithm · Two-cluster classification virial error | convergence order: 0.86 -> 2.01 | Improved |
| add EQN preconditioning with more walkers (8 to 16) at extra cost | EQN sampler · MNIST two-class (7 vs 9) | per-walker overhead: 40% -> 50% | Hurt |
| use EQN on MNIST with scale-invariant output biases left free | EQN sampler · MNIST two-class | IAT reduction | No clear effect |
Closest papers by meaning
- Microcanonical Langevin Ensembles: Advancing the Sampling of Bayesian Neural Networks (ICLR 2025 · similarity 0.88)
- Parameter Expanded Stochastic Gradient Markov Chain Monte Carlo (ICLR 2025 · similarity 0.88)
- Scalable Bayesian Learning with posteriors (ICLR 2025 · similarity 0.88)
- Stop the Sampler! Classifier-Based Adaptive Stopping for Sampling Kernels (NeurIPS 2026 · similarity 0.87)
- A Study of Bayesian Neural Network Surrogates for Bayesian Optimization (ICLR 2024 · similarity 0.86)
- Inflationary Flows: Calibrated Bayesian Inference with Diffusion-Based Models (NeurIPS 2024 · similarity 0.86)
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.
