Research map / Label-efficient and robust learning / Recognition and long-tail
Just Train Twice: Improving Group Robustness without Training Group Information
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 |
|---|---|---|---|
| Upweight examples misclassified by an early-stopped ERM identification model | ERM · Waterbirds | worst-group accuracy: 72.6% -> 86.7% | Helped |
| Upweight examples misclassified by an early-stopped ERM identification model | ERM · CelebA | worst-group accuracy: 47.2% -> 81.1% | Helped |
| Upweight examples misclassified by an early-stopped ERM identification model | ERM · MultiNLI | worst-group accuracy: 67.9% -> 72.6% | Helped |
| Upweight examples misclassified by an early-stopped ERM identification model | ERM · CivilComments-WILDS | worst-group accuracy: 57.4% -> 69.3% | Helped |
| Remove waterbirds on water backgrounds from the error set (ablation) | JTT on Waterbirds · Waterbirds | worst-group accuracy: 86.7% -> 80.7% | Hurt |
| Replace error-set examples with random examples from the same group (ablation) | JTT on Waterbirds · Waterbirds | worst-group accuracy: 86.7% -> 86% | No clear effect |
| Recompute the error set every 10 epochs instead of keeping it fixed (ablation) | JTT on Waterbirds · Waterbirds | worst-group accuracy: 88.6% -> 77.1% | Hurt |
| Recompute the error set every 20 epochs instead of keeping it fixed (ablation) | JTT on Waterbirds · Waterbirds | worst-group accuracy: 88.6% -> 72.1% | Hurt |
| Recompute the error set every 30 epochs instead of keeping it fixed (ablation) | JTT on Waterbirds · Waterbirds | worst-group accuracy: 88.6% -> 86.8% | No clear effect |
| Tune hyperparameters for average rather than worst-group validation accuracy (ablation) | JTT · Waterbirds | worst-group test accuracy: 62.5% -> 86.7% | Hurt |
| Tune hyperparameters for average rather than worst-group validation accuracy (ablation) | JTT · CelebA | worst-group test accuracy: 40.6% -> 81.1% | Hurt |
| Remove either the y=a or y≠a examples from the error set (ablation) | JTT · Waterbirds and CelebA | worst-group accuracy | Hurt |
Closest papers by meaning
- The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 2024 · similarity 0.87)
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.87)
- Rank-Constrained Adaptation for Reliable Real-World Performance (NeurIPS 2026 · similarity 0.87)
- Project-Probe-Aggregate: Efficient Fine-Tuning for Group Robustness (CVPR 2025 · similarity 0.87)
- Improving Subgroup Robustness via Data Selection (NeurIPS 2024 · similarity 0.87)
- ExMap: Leveraging Explainability Heatmaps for Unsupervised Group Robustness to Spurious Correlations (CVPR 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.