Research map / Label-efficient and robust learning / Noisy labels
R-divergence for Estimating Model-oriented Distribution Discrepancy
Zhilin Zhao, Longbing Cao
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 |
|---|---|---|---|
| R-Div instead of H-Div (hypothesis learned on mixed data, risk on each dataset) | distribution discrepancy test · HIGGS average | test power: 0.847 -> 1.000 | Improved |
| R-Div instead of H-Div at N=1000 | distribution discrepancy test · HIGGS | test power: 0.240 -> 1.000 | Improved |
| R-Div instead of H-Div | distribution discrepancy test · MNIST vs Colored-MNIST | test power: 0.951 -> 1.000 | Improved |
| R-Div instead of H-Div | distribution discrepancy test · PACS six domain pairs average | test power: 0.714 -> 0.849 | Improved |
| R-Div instead of H-Div on hardest domain pair C+P | distribution discrepancy test · PACS C+P | test power: 0.448 -> 0.527 | Improved |
| R-Div instead of H-Div on SENet at N=5000 | distribution discrepancy test · CIFAR10 vs corrupted CIFAR10 | test power: 0.478 -> 0.790 | Improved |
| R-Div instead of H-Div on ViT at N=1000 | distribution discrepancy test · CIFAR10 vs corrupted CIFAR10 | test power: 0.119 -> 0.211 | Improved |
| Increase sample size to N=10000 | R-Div and H-Div · CIFAR10 vs corrupted CIFAR10 | test power: 1.000 / 1.000 | No clear effect |
| R-Div instead of H-Div on MNIST | distribution discrepancy test · MNIST | test power: 1.000 -> 1.000 | No clear effect |
| R-Div instead of H-Div at 1% corruption | distribution discrepancy test · CIFAR10 Snow corruption | test power: 0.096 -> 0.189 | Improved |
| R-Div instead of H-Div at 0.5% corruption (mild) | distribution discrepancy test · CIFAR10 and ImageNet corruptions | test power: 0.000 / 0.000 | No clear effect |
| Mixed-data minimum hypothesis (R-Div) vs per-dataset hypotheses (H-Div) | discrepancy estimator · MNIST N=200 | risk gap: 7.1813 -> 44.4703 | Improved |
Closest papers by meaning
- Revealing Distribution Discrepancy by Sampling Transfer in Unlabeled Data (NeurIPS 2024 · similarity 0.88)
- Neural Estimation of Statistical Divergences (JMLR 2022 · similarity 0.87)
- On $f$-Divergence Principled Domain Adaptation: An Improved Framework (NeurIPS 2024 · similarity 0.84)
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · similarity 0.84)
- Reproduce, Replicate, Reevaluate. The Long but Safe Way to Extend Machine Learning Methods (AAAI 2024 · similarity 0.84)
- "What is Different Between These Datasets?" A Framework for Explaining Data Distribution Shifts (NeurIPS 2026 · 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.
