Research map / Label-efficient and robust learning / Recognition and long-tail
Examining and Combating Spurious Features under Distribution Shift
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 |
|---|---|---|---|
| Use instance-level weights within groups in addition to group-level weights (GC-DRO instead of group DRO) | CelebA with imperfect partitions · CelebA imperfect partition | robust accuracy: 45.97 $\pm$ 1.73 -> 82.85 $\pm$ 1.54 | Helped |
| Use instance-level weights within groups in addition to group-level weights (GC-DRO instead of group DRO) | MNLI with imperfect partitions · MNLI imperfect partition | robust accuracy: 70.34 $\pm$ 2.19 -> 75.32 $\pm$ 0.93 | Helped |
| Use instance-level weights within groups in addition to group-level weights (GC-DRO instead of group DRO) | FDCL18 with imperfect partitions · FDCL18 imperfect partition | robust accuracy: 36.24 $\pm$ 3.80 -> 48.42 $\pm$ 6.72 | Helped |
| Use instance-level weights within groups in addition to group-level weights (GC-DRO instead of group DRO) | FDCL18 with unsupervised-cluster partitions · FDCL18 unsupervised clustering | robust accuracy: 34.48 $\pm$ 4.69 -> 45.06 $\pm$ 6.77 | Helped |
| Use GC-DRO rather than ERM | CelebA with clean partitions · CelebA clean partition | robust accuracy: 40.14 $\pm$ 0.99 -> 88.75 $\pm$ 0.82 | Helped |
| Use GC-DRO rather than group DRO (EG) | MNLI with clean partitions · MNLI clean partition | robust accuracy: 77.88 $\pm$ 1.36 -> 77.82 $\pm$ 1.45 | No clear effect |
| Use GC-DRO rather than ERM | FDCL18 with clean partitions · FDCL18 clean partition | robust accuracy: 34.30 $\pm$ 1.83 -> 57.28 $\pm$ 2.71 | Helped |
| Use GC-DRO rather than the best average-accuracy baseline | CelebA with imperfect partitions · CelebA imperfect partition | average accuracy: 95.92 $\pm$ 0.05 -> 89.32 $\pm$ 2.21 | Hurt |
| Use GC-DRO rather than the best average-accuracy baseline | FDCL18 with imperfect partitions · FDCL18 imperfect partition | average accuracy: 80.66 $\pm$ 0.52 -> 72.02 $\pm$ 2.96 | Hurt |
| Vary alpha and beta; under clean partitions, use larger beta and smaller alpha (ablation) | GC-DRO on MNLI · MNLI clean and imperfect partitions | robust accuracy | Mixed |
Closest papers by meaning
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.88)
- MixMax: Distributional Robustness in Function Space via Optimal Data Mixtures (ICLR 2025 · similarity 0.88)
- Sufficient Invariant Learning for Distribution Shift (CVPR 2025 · similarity 0.88)
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.88)
- Improving Group Robustness on Spurious Correlation Requires Preciser Group Inference (ICML 2024 · similarity 0.87)
- The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 2024 · similarity 0.87)
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.