Research map / Probabilistic and scientific modeling / Variational and generative sampling
Understanding Deep Generative Models With Generalized Empirical Likelihoods
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 GEL on Pool3 features with KGEL (kernelized moment condition with 1024 witness points) | mode dropping detection on CIFAR-10 · CIFAR-10 mode dropping | Hellinger distance to oracle probabilities: 0.2732 vs 0.6137 (Improved Recall) at 8 missing modes | Helped |
| Use KGEL instead of improved recall/coverage for mode-imbalance estimation | mode imbalance experiment on CIFAR-10 · CIFAR-10 mode imbalance | Hellinger distance: 0.1178 vs 0.3016 (Improved Recall) at probability 0.1 | Helped |
| Add label information (product kernel with delta label kernel) to KGEL to detect improper label conditioning | CIFAR-10 label corruption detection · CIFAR-10 with 10-60% label errors | Precision-recall (AUC) | Helped |
| Use exponential tilting vs empirical likelihood objective in KGEL2 on ImageNet | two-sample GEL on ImageNet models · ImageNet 256x256 | 2^D score | Mixed |
| Stronger truncation (tau from 1.0 to 0.2) for BigGAN-deep (ablation) | BigGAN-deep on ImageNet · ImageNet 256x256 KGEL EL | 2^D score: 1.735 at tau=1.0 to 3.965 at tau=0.2 | Hurt |
| Increase classifier guidance scale of ADM beyond 1.0 (ablation) | ADM-G on ImageNet · ImageNet 256x256 KGEL EL | 2^D score: 1.786 at scale 1.0 to 3.219 at scale 10.0 | Hurt |
| Increase number of witness points from 64 to 1024 in KGEL mode dropping (ablation) | KGEL on CIFAR-10 Pool3 mode dropping · CIFAR-10 mode dropping | Hellinger distance: 0.1466 with 1024 vs 0.1891 with 64 (GEL Pool3 2 missing modes) | Helped |
Closest papers by meaning
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.88)
- A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models (ICML 2025 · similarity 0.87)
- Glauber Generative Model: Discrete Diffusion Models via Binary Classification (ICLR 2025 · similarity 0.86)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.86)
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 2024 · similarity 0.86)
- An Interpretable Evaluation of Entropy-based Novelty of Generative Models (ICML 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.