Research map / Probabilistic and scientific modeling / Variational and generative sampling
MCMC Should Mix: Learning Energy-Based Model with Neural Transport Latent Space MCMC
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Exponentially tilt the Glow reference with an energy correction and train using neural-transport MCMC | Glow · MNIST | FID: 66.04 -> 21.32 | Helped |
| Exponentially tilt the Glow reference with an energy correction and train using neural-transport MCMC | Glow · SVHN | FID: 94.23 -> 48.01 | Helped |
| Exponentially tilt the Glow reference with an energy correction and train using neural-transport MCMC | Glow · CelebA | FID: 59.35 -> 46.38 | Helped |
| Exponentially tilt the Glow reference with an energy correction and train using neural-transport MCMC | Glow · CIFAR-10 | FID: 90.08 -> 78.12 | Helped |
| Train the correction with NCE rather than neural-transport MCMC | Glow-backed EBM · MNIST | FID: 21.32 -> 36.52 | Hurt |
| Train the correction with NCE rather than neural-transport MCMC | Glow-backed EBM · SVHN | FID: 48.01 -> 79.84 | Hurt |
| Train the correction with NCE rather than neural-transport MCMC | Glow-backed EBM · CelebA | FID: 46.38 -> 51.73 | Hurt |
| Use neural-transport HMC in latent space instead of HMC in data space | EBM sampling · SVHN | Gelman-Rubin statistic (mean R-hat) | Helped |
| Use neural-transport HMC in latent space instead of over-damped Langevin in latent space | EBM sampling · SVHN | autocorrelation decay lag: within Δt = 200 steps; Δt > 1,000 steps | Helped |
| Increase the flow backbone from Small to Medium (ablation) | NT-EBM · SVHN | FID: 74.77 -> 48.01 | Helped |
| Increase the flow backbone from Medium to Large (ablation) | NT-EBM · SVHN | FID: 48.01 -> 43.82 | Helped |
| Increase Glow backbone size from Small to Largest (ablation) | Glow (MLE) · SVHN | FID: 110.55 -> 86.18 | Helped |
Closest papers by meaning
- Learning Energy-Based Prior Model with Diffusion-Amortized MCMC (NeurIPS 2023 · similarity 0.89)
- Learning Latent Space Hierarchical EBM Diffusion Models (ICML 2024 · similarity 0.89)
- Energy-guided Entropic Neural Optimal Transport (ICLR 2024 · similarity 0.86)
- Joint Learning of Energy-based Models and their Partition Function (ICML 2025 · similarity 0.86)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.85)
- Energy-Based Modelling for Discrete and Mixed Data via Heat Equations on Structured Spaces (NeurIPS 2024 · similarity 0.85)
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.