Research map / Learning theory and algorithms / Optimal transport
Performance Scaling via Optimal Transport: Enabling Data Selection from Partially Revealed Sources
Feiyang Kang, Hoang Anh Just, Anit Kumar Sahu, Ruoxi Jia
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 |
|---|---|---|---|
| use projektor/PQ instead of Linear | performance predictor · CIFAR-10 extrapolation | test MAE: 6.04 -> 1.26 | Improved |
| use projektor/PQ instead of Quadratic | performance predictor · CIFAR-10 extrapolation | test MAE: 5.22 -> 1.26 | Improved |
| use projektor/PQ instead of Rational | performance predictor · ImageNet100 extrapolation | test MAE: 2.27 -> 0.38 | Improved |
| use projektor/PQ instead of Quadratic | performance predictor · IMDB extrapolation | test MAE: 2.74 -> 1.16 | Improved |
| use projektor/PQ instead of Linear | performance predictor · MNIST extrapolation | test MAE: 28.03 -> 7.27 | Improved |
| use projektor/CS instead of projektor/PQ (ablation) | performance predictor · MNIST extrapolation | test MAE: 7.27 -> 4.26 | Improved |
| use PQ (b,c functions of mixing ratio) instead of CS (ablation) | projektor · CIFAR-10, ImageNet100, IMDB extrapolation | test MAE: 1.95 -> 1.26 | Improved |
| PQ gives lower training MAE than CS (ablation) | projektor · IMDB training fit | train MAE: 0.90 -> 0.25 | Improved |
| add OT distance as feature to quadratic-style predictors (projektor/PQ vs Pseudo-Quadratic) (ablation) | performance predictor · CIFAR-10 extrapolation | test MAE: 2.90 -> 1.26 | Improved |
| fit with more compositions: CS needs about 15 trainings, PQ about 25 to converge | projektor · CIFAR-10 extrapolation | MAE vs number of trainings: 15 vs 25 trainings | Improved |
| use projektor for mislabeled sources performance projection to 2K-10K from 1K | projektor/PQ · CIFAR-10 mislabeled sources | MAE: below 2% | Improved |
| use projektor for unlabeled sources projection to larger scales | projektor · CIFAR-10 unlabeled sources | projection error: below 10% | Improved |
Closest papers by meaning
- Adaptive Data Optimization: Dynamic Sample Selection with Scaling Laws (ICLR 2025 · similarity 0.86)
- Scaling Laws for Optimal Data Mixtures (NeurIPS 2025 · similarity 0.85)
- DsDm: Model-Aware Dataset Selection with Datamodels (ICML 2024 · similarity 0.85)
- Data Mixture Optimization: A Multi-fidelity Multi-scale Bayesian Framework (NeurIPS 2025 · similarity 0.85)
- DataDecide: How to Predict Best Pretraining Data with Small Experiments (ICML 2025 · similarity 0.85)
- Compute-Constrained Data Selection (ICLR 2025 · 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.
