Research map / Label-efficient and robust learning / Recognition and long-tail
Change is Hard: A Closer Look at Subpopulation 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 two-stage classifier retraining (DFR, CRT, ReWeightCRT) instead of ERM | subpopulation shift methods · 12 datasets, attributes unknown in train and val | worst-group accuracy (avg): 66.2/66.1/65.4 vs 59.9 | Helped |
| balanced classifier learning (resampling/reweighting last stage) (ablation) | ResNet-50 representation for spurious correlation and class imbalance shifts · SC and CI dataset groups | worst-group accuracy relative to ERM: +8.1 (SC), +11.9 (CI) | Helped |
| balanced representation learning (ablation) | attribute imbalance shift · AI dataset group | worst-group accuracy relative to ERM: +1.1 | Helped |
| balanced classifier learning (ablation) | attribute imbalance shift · AI dataset group | worst-group accuracy relative to ERM: +0.0 | No clear effect |
| stratified balanced representation or classifier learning (ablation) | attribute generalization shift · AG dataset group | worst-group accuracy relative to ERM: -0.4 (representation), -0.4 (classifier) | Hurt |
| select model by worst-class validation accuracy (no group labels) instead of oracle (ablation) | model selection · 6 datasets, attributes unknown | worst-group test accuracy gap to oracle: -1.8 avg | No clear effect |
| select model by overall validation accuracy instead of worst-group (ablation) | model selection · 6 datasets, attributes unknown | worst-group test accuracy gap to oracle: -15.0 avg | Hurt |
| algorithms that use group-based methods improve worst-group accuracy on SC and CI shifts but not AI or AG | SOTA subpopulation shift algorithms · Waterbirds, CelebA, etc. grouped by shift type | worst-group accuracy vs ERM | Mixed |
| training with group-DRO when training attributes are available (vs unknown) (ablation) | GroupDRO · attribute availability comparison | worst-group accuracy | Helped |
Closest papers by meaning
- Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift (ICML 2025 · similarity 0.88)
- Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions (ICLR 2025 · similarity 0.87)
- Learning With Multi-Group Guarantees For Clusterable Subpopulations (ICML 2025 · similarity 0.86)
- CODA: A Correlation-Oriented Disentanglement and Augmentation Modeling Scheme for Better Resisting Subpopulation Shifts (NeurIPS 2024 · similarity 0.86)
- Evaluating Model Bias Requires Characterizing its Mistakes (ICML 2024 · similarity 0.86)
- "Who experiences large model decay and why?" A Hierarchical Framework for Diagnosing Heterogeneous Performance Drift (ICML 2025 · 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.