Research map / Architectures and efficiency / RNNs and normalization
Towards Last-layer Retraining for Group Robustness with Fewer Annotations
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 |
|---|---|---|---|
| Class-balance the last-layer retraining data rather than leave it unbalanced (ablation) | Last-layer retraining · CelebA | worst-group accuracy: 41.9 ± 1.4 -> 73.7 ± 2.8 | Helped |
| Class-balance the last-layer retraining data rather than leave it unbalanced (ablation) | Last-layer retraining · CivilComments | worst-group accuracy: 57.6 ± 4.2 -> 80.4 ± 0.8 | Helped |
| Class-balance the last-layer retraining data rather than leave it unbalanced (ablation) | Last-layer retraining · Waterbirds | worst-group accuracy: 88.0 ± 0.8 -> 92.6 ± 0.8 | Helped |
| Class-balance the last-layer retraining data rather than leave it unbalanced (ablation) | Last-layer retraining · MultiNLI | worst-group accuracy: 64.6 ± 1.0 -> 64.7 ± 1.1 | No clear effect |
| Select disagreement points between ERM and early-stopped models instead of misclassified points (ablation) | SELF · CivilComments | worst-group accuracy: 62.7 ± 4.6 -> 79.1 ± 2.1 | Helped |
| Select disagreement points between ERM and early-stopped models instead of misclassified points (ablation) | SELF · CelebA | worst-group accuracy: 83.0 ± 6.1 -> 83.9 ± 0.9 | Helped |
| Select disagreement points between ERM and early-stopped models instead of misclassified points (ablation) | SELF · MultiNLI | worst-group accuracy: 72.2 ± 2.2 -> 70.7 ± 2.5 | Hurt |
| Select disagreement points between ERM and early-stopped models instead of misclassified points (ablation) | SELF · Waterbirds | worst-group accuracy: 92.6 ± 0.8 -> 93.0 ± 0.3 | No clear effect |
| Use early-stop disagreement rather than dropout disagreement (ablation) | SELF · CivilComments | worst-group accuracy: 69.9 ± 5.2 -> 79.1 ± 2.1 | Helped |
| Use early-stop disagreement rather than dropout disagreement (ablation) | SELF · MultiNLI | worst-group accuracy: 68.7 ± 3.4 -> 70.7 ± 2.5 | Helped |
| Use early-stop disagreement rather than dropout disagreement (ablation) | SELF · CelebA | worst-group accuracy: 85.7 ± 1.6 -> 83.9 ± 0.9 | Hurt |
| Use early-stop disagreement rather than dropout disagreement (ablation) | SELF · Waterbirds | worst-group accuracy: 92.3 ± 0.5 -> 93.0 ± 0.3 | Mixed |
Closest papers by meaning
- The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 2024 · similarity 0.91)
- Enhancing Robustness of Last Layer Two-Stage Fair Model Corrections (NeurIPS 2024 · similarity 0.89)
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.87)
- Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions (ICLR 2025 · similarity 0.87)
- Calibrating Multi-modal Representations: A Pursuit of Group Robustness without Annotations (CVPR 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.