Research map / Label-efficient and robust learning / Recognition and long-tail
Distributionally Robust Optimization with Probabilistic Group
Soumya Suvra Ghosal, Yixuan Li
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 probabilistic group membership rather than hard group labels (ablation) | PG-DRO versus G-DRO · Waterbirds and CelebA | worst-group accuracy: Waterbirds: 88.0 -> 91.0; CelebA: 88.6 -> 90.0 | Improved |
| use probabilistic group membership rather than hard group labels with 5% validation labels (ablation) | PG-DRO versus G-DRO · Waterbirds and CelebA | worst-group accuracy: Waterbirds: 87.5 -> 89.2; CelebA: 87.5 -> 89.4 | Improved |
| use probabilistic group membership with semi-supervised pseudo-labeling (ablation) | PG-DRO versus G-DRO · Waterbirds | worst-group accuracy: 89.0 -> 91.0 | Improved |
| use probabilistic group membership with semi-supervised pseudo-labeling (ablation) | PG-DRO versus G-DRO · CelebA | worst-group accuracy: 89.8 -> 89.8 | No clear effect |
| use probabilistic group membership rather than hard group labels with zero-shot pseudo-labeling (ablation) | PG-DRO versus G-DRO · Waterbirds and CelebA | worst-group accuracy: Waterbirds: 87.6 -> 89.4; CelebA: 87.8 -> 88.8 | Improved |
| use CLIP zero-shot group probabilities rather than supervised pseudo-labeling (ablation) | PG-DRO · Waterbirds and CelebA | worst-group accuracy: Waterbirds: 89.2 -> 89.4; CelebA: 89.4 -> 88.8 | Mixed |
| tune the generalization adjustment parameter to C=2 rather than use C=0 (ablation) | PG-DRO · Waterbirds | worst-group accuracy: +5.1% | Improved |
| reduce group-annotated validation data to 5% (ablation) | PG-DRO · Waterbirds and CelebA | worst-group accuracy: Waterbirds: 91.0 -> 89.2; CelebA: 90.0 -> 89.4 | Mixed |
| use PG-DRO rather than SSA | image classification · Waterbirds and CelebA | worst-group accuracy: Waterbirds: 89.0 -> 91.0; CelebA: 89.8 -> 90.0 | Mixed |
| use PG-DRO rather than SSA | natural language processing · MultiNLI and CivilComments-WILDS | worst-group accuracy: MultiNLI: 76.6 -> 78.9; CivilComments-WILDS: 69.9 -> 73.6 | Improved |
| use PG-DRO rather than G-DRO | robust training with group probabilities · CelebA | worst-group accuracy: 88.7 -> 90.0 | Improved |
Closest papers by meaning
- MixMax: Distributional Robustness in Function Space via Optimal Data Mixtures (ICLR 2025 · similarity 0.88)
- Efficient Algorithms for Empirical Group Distributionally Robust Optimization and Beyond (ICML 2024 · similarity 0.88)
- Beyond Minimax Rates in Group Distributionally Robust Optimization via a Novel Notion of Sparsity (ICML 2025 · similarity 0.86)
- Group Distributionally Robust Optimization with Flexible Sample Queries (NeurIPS 2026 · similarity 0.86)
- Structure-informed Risk Minimization for Robust Ensemble Learning (ICML 2025 · similarity 0.85)
- Project-Probe-Aggregate: Efficient Fine-Tuning for Group Robustness (CVPR 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.
