Research map / Probabilistic and scientific modeling / Variational and generative sampling
Learning the Stein Discrepancy for Training and Evaluating Energy-Based Models without Sampling
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 |
|---|---|---|---|
| replace linear-time kernel Stein tests with a learned neural critic | LSD goodness-of-fit test · Gaussian-Bernoulli RBMs as dimension increases | rejection rate | Helped |
| use a learned critic rather than the quadratic-time KSD | LSD goodness-of-fit test · High-dimensional Gaussian-Bernoulli RBMs | rejection rate | No clear effect |
| use the MMD test requiring model samples instead of a Stein test | Goodness-of-fit testing · Gaussian-Bernoulli RBMs as dimension increases | rejection rate | Hurt |
| use the learned Stein discrepancy instead of kernel-based KSD estimates | Gaussian-Bernoulli RBM model evaluation · Dimensions 50, 100, and 200 with perturbed model weights | sensitivity to model perturbations | Helped |
| use LSD instead of score matching as a flow-evaluation score | Glow · MNIST test set checkpoints | correlation with likelihood | Helped |
| train the unnormalized model with LSD rather than NCE, CNCE, or score matching | Linear ICA · Dimensions above 20 | test log-likelihood | Helped |
| use score matching rather than LSD for training | Linear ICA · Dimensions above 30 | training stability | Hurt |
| use 1 rather than 5 critic steps per model update (ablation) | LSD training · Linear ICA, all tested dimensions | test performance | Helped |
| increase Langevin noise scale from .01 to .1 (ablation) | SGLD sampling from trained deep EBMs · MNIST and FashionMNIST sample visualization | sample diversity | Helped |
| increase SGLD sampling from 30 to 1000 steps (ablation) | SGLD sampling from trained deep EBMs · MNIST and FashionMNIST sample visualization | visual sample clarity | Helped |
| select critics using validation mean minus standard deviation rather than validation mean alone (ablation) | LSD model comparison · 10-dimensional Gaussian data with an overfitting critic | critic model selection under overfitting | Helped |
Closest papers by meaning
- Autoregressive Energy Machines (ICML 2019 · similarity 0.88)
- Approximate Stein Classes for Truncated Density Estimation (ICML 2023 · similarity 0.87)
- Learning normalized image densities via dual score matching (NeurIPS 2025 · similarity 0.85)
- The Polynomial Stein Discrepancy for Assessing Moment Convergence (ICML 2025 · similarity 0.85)
- Energy-Based Modelling for Discrete and Mixed Data via Heat Equations on Structured Spaces (NeurIPS 2024 · similarity 0.84)
- Generative Marginalization Models (ICML 2024 · similarity 0.84)
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.