Research map / Label-efficient and robust learning / Recognition and long-tail
Correct-N-Contrast: a Contrastive Approach for Improving Robustness to Spurious Correlations
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 |
|---|---|---|---|
| CNC two-stage contrastive learning with ERM-inferred pseudo groups instead of JTT upweighting | worst-group accuracy without group labels · CMNIST*, Waterbirds, CelebA, CivilComments-WILDS | worst-group accuracy: CelebA +7.3 vs JTT; CivilComments -0.4 vs JTT | Mixed |
| Replace Stage 1 ERM-inferred groups with true group labels (CNC*) (ablation) | CNC · Waterbirds and CelebA | worst-group accuracy: Waterbirds 90.1 vs GDRO 89.9; CelebA 89.2 vs GDRO 88.9 | Helped |
| Negatives sampled only from different-class samples (removing same-ERM-prediction criterion) (ablation) | CNC sampling · Waterbirds | worst-group accuracy: 82.2 vs 88.5 default | Hurt |
| Negatives sampled only from same ERM prediction (removing different-class criterion) (ablation) | CNC sampling · CelebA | worst-group accuracy: 75.0 vs 88.8 default | Hurt |
| Positives sampled only from same class (removing different-ERM-prediction criterion) (ablation) | CNC sampling · CelebA | worst-group accuracy: 74.4 vs 88.8 default | Hurt |
| Class-only supervised contrastive sampling (no ERM-based hard positives/negatives) (ablation) | CNC sampling · Waterbirds and CelebA | worst-group accuracy: Waterbirds 62.3 vs 88.5; CelebA 61.5 vs 88.8 | Hurt |
| Stage 1 label noise added to ERM predictions (p=0.25) (ablation) | CNC vs JTT · CMNIST* | worst-group accuracy: CNC 53.2 vs JTT 43.6 at p=0.25; CNC 80.6 vs JTT 78.4 at p=0 | Mixed |
| Cluster ERM hidden-layer outputs into C clusters for pseudo-groups instead of argmax predictions (ablation) | CNC Stage 1 · Stage 1 pseudo-group inference | worst-group accuracy | Helped |
Closest papers by meaning
- Improving Group Robustness on Spurious Correlation Requires Preciser Group Inference (ICML 2024 · similarity 0.89)
- Examining and Combating Spurious Features under Distribution Shift (ICML 2021 · similarity 0.88)
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.87)
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.86)
- The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 2024 · similarity 0.86)
- Prompting is a Double-Edged Sword: Improving Worst-Group Robustness of Foundation Models (ICML 2024 · 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.