Research map / Label-efficient and robust learning / Recognition and long-tail
SEBRA : Debiasing through Self-Guided Bias Ranking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Sebra, an unsupervised self-guided method that ranks training samples by spuriosity via a modulated ERM training trajectory, then uses the ranking in a contrastive debiasing loss.
Problem
Debiasing neural networks against spurious correlations without human supervision, using fine-grained spuriosity rankings of samples rather than binary biased/unbiased partitions.
Approach
Exploits a hardness-spuriosity symmetry in ERM: samples learned earlier carry more spurious cues. Sebra steers ERM with selection variables v and upweighting variables u so samples are removed from training once learned, yielding a rank order from high to low spuriosity without human supervision; the ranking then drives pair selection in a supervised contrastive loss.
Key results
- Kendall's tau of 0.85 on UrbanCars for Sebra ranking vs 0.40 for Spuriosity Ranking (supervised) and 0.12 for ERM-based ranking.
- On BAR, Sebra reaches 75.36% test accuracy vs 69.88% for Debian, the best unsupervised baseline.
- On CelebA, Sebra achieves 65.3% worst-group accuracy vs 41.1% for Debian (unsupervised baseline).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Sebra's self-guided ranking instead of Spuriosity Ranking | spuriosity ranking · UrbanCars | Kendall's τ: 0.40 -> 0.85 | Helped |
| Use Sebra's self-guided ranking instead of Spuriosity Ranking | spuriosity ranking · CelebA | Kendall's τ: 0.38 -> 0.69 | Helped |
| Use Sebra's self-guided ranking instead of Spuriosity Ranking | spuriosity ranking · BAR | Performance Disparity (PD): 28.88 -> 32.32 | Helped |
| Use Sebra instead of Debian | debiasing · BAR | Test Acc.: 69.88 (2.92) -> 75.36 (2.23) | Helped |
| Use Sebra instead of Group DRO | debiasing · CelebA | WG Acc.: 37.9 (1.6) -> 65.3 (4.1) | Helped |
| Use Sebra instead of JTT | debiasing · MultiNLI | WG. Acc: 69.1 -> 72.3 | Helped |
| Remove v_i selection and u_i weighting (ablation) | Sebra · UrbanCars | Kendall's τ: 0.85 -> 0.12 | Hurt |
| Add v_i selection (ablation) | Sebra · UrbanCars | Kendall's τ: 0.12 -> 0.79 | Helped |
| Remove u_i weighting (ablation) | Sebra · UrbanCars | Kendall's τ: 0.85 -> 0.79 | Hurt |
| Increase β (ablation) | Sebra ranking · ranking quality | Kendall's τ | Hurt |
Surprising findings
- Sebra's ranking quality decreases almost linearly as beta increases, favoring small beta values.
- ERM alone ranks data poorly (Kendall's tau 0.12), with the ranking gain coming from the selection and upweighting terms.
Limitations
- Ranking quality is sensitive to the beta hyperparameter.
- Contrastive debiasing is considerably more expensive than the ranking step.
- Unsupervised ranking metrics are lacking, so evaluation relies on coarse bias annotations or proxy metrics.
Experimental setup
Datasets: UrbanCars CelebA BAR ImageNet-1K MultiNLI Living17 Landbirds
Benchmarks: UrbanCars CelebA BAR ImageNet-1K MultiNLI
Metrics: I.D. Acc worst-group accuracy Avg GAP Kendall's tau Performance Disparity BAR test accuracy
Baselines: Group DRO ERM LfF JTT Debian DFR Spuriosity Ranking LLE
Builds on: Learning from Failure Just Train Twice Deep Feature Reweighting Spuriosity Ranking Supervised contrastive learning
Closest papers by meaning
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.87)
- Mitigating Spurious Correlations via Disagreement Probability (NeurIPS 2024 · similarity 0.87)
- A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective (NeurIPS 2024 · similarity 0.87)
- Limits to scalable evaluation at the frontier: LLM as judge won’t beat twice the data (ICLR 2025 · similarity 0.87)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.86)
- Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · 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.