Research map / Probabilistic and scientific modeling / Variational and generative sampling
On the Anatomy of MCMC-Based Maximum Likelihood Learning of Energy-Based Models
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 |
|---|---|---|---|
| Use noise initialization instead of informative initialization during training | ConvNet potential trained with non-convergent ML · MNIST, Oxford Flowers 102, CelebA, CIFAR-10 | sample realism and diversity from noise | Helped |
| Allow MCMC training samples to remain non-convergent instead of requiring steady-state convergence | ConvNet potential trained with non-convergent ML · Image synthesis | short-run sample realism | No clear effect |
| Use non-convergent rather than convergent ML | ConvNet potential · Image synthesis | steady-state sample realism | Hurt |
| Tune Langevin noise to support convergence | ConvNet potential trained with persistent initialization · Oxford Flowers 102 | steady-state sample realism: ε = 0.015 | Helped |
| Add non-local layers | Convergent ML · Oxford Flowers 102 | synthesis realism | Helped |
| Use informative initialization rather than noise initialization | ConvNet potential trained for non-convergent ML · Image synthesis from noise | sample diversity | Hurt |
| Use informative initialization rather than noise initialization | ConvNet potential trained for convergent ML · Image synthesis | MCMC steps needed for convergence: L ≈ 20,000; L = 500 | Helped |
| Increase Langevin noise from too-low to appropriately tuned | ML learning on 2D toy distributions · 2D toy distributions | learned density match to ground truth: ε = 0.01; ε = 0.125 | Helped |
| Use Adam instead of SGD | ConvNet potential trained for convergent ML · Image synthesis | steady-state realism | Hurt |
| Use Adam instead of SGD | ConvNet potential trained for non-convergent ML · Image synthesis | training speed and image quality | Helped |
| Use a 3×3 stride-1 first convolution | ConvNet potential · Image synthesis | checkerboard and other artifacts | Helped |
| Add regularization or normalization techniques (ablation) | ML learning of ConvNet potentials · Image synthesis and training stability | training stability | No clear effect |
Closest papers by meaning
- Learning Non-Convergent Non-Persistent Short-Run MCMC Toward Energy-Based Model (NeurIPS 2019 · similarity 0.91)
- Learning Energy-based Model via Dual-MCMC Teaching (NeurIPS 2023 · similarity 0.91)
- AutoStep: Locally adaptive involutive MCMC (ICML 2025 · similarity 0.88)
- Parameter Expanded Stochastic Gradient Markov Chain Monte Carlo (ICLR 2025 · similarity 0.88)
- Microcanonical Langevin Ensembles: Advancing the Sampling of Bayesian Neural Networks (ICLR 2025 · similarity 0.88)
- Learning to Explore for Stochastic Gradient MCMC (ICML 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.