Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Debiased Classifier with Biased Committee
Nayeong Kim, Sehyun Hwang, Sungsoo Ahn, Jaesik Park, Suha Kwak
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| up-weight samples misclassified by a single biased classifier (vs plain ERM) (ablation) | SSL+ERM on CelebA HairColor · CelebA HairColor | Worst-group accuracy: 38.5 -> 64.1 | Improved |
| use a committee of 40 biased classifiers instead of a single one (ablation) | LWBC · CelebA HairColor | Worst-group accuracy: 64.1 -> 81.3 | Improved |
| add knowledge distillation from main classifier to the committee (ablation) | LWBC · CelebA HairColor | Worst-group accuracy: 81.3 -> 85.5 | Improved |
| add knowledge distillation to a single-classifier variant (ablation) | LWBC single member · CelebA HairColor | Worst-group accuracy: 72.7 -> 77.9 | Improved |
| train committee members on bootstrapped random subsets instead of entire data (ablation) | LWBC · CelebA HairColor | Worst-group accuracy: 78.0 -> 81.3 | Improved |
| replace JTT weighting with LWBC weight function (ablation) | SSL single classifier · CelebA HairColor | Worst-group accuracy: 57.7 -> 64.1 | Improved |
| use self-supervised features (vs none) (ablation) | LWBC · CelebA HairColor | Worst-group accuracy: 80.8 -> 85.5 | Improved |
| bootstrapped committee vs deep ensemble (ablation) | LWBC · CelebA HairColor | Worst-group accuracy: 77.2 -> 85.5 | Improved |
| increase committee size beyond 40 (ablation) | LWBC · CelebA HairColor | Worst-group accuracy | No clear effect |
| self-supervised representation instead of supervised ERM | ERM · CelebA HairColor | Conflicting accuracy: 52.52 -> 66.79 | Improved |
| add LWBC on top of SSL+ERM | SSL+ERM · BAR | Conflicting accuracy: 60.88 -> 62.03 | Improved |
| add LWBC on top of SSL+ERM | SSL+ERM · ImageNet-9 | Unbiased accuracy: 93.18 -> 93.04 | No clear effect |
Closest papers by meaning
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.86)
- Improving Subgroup Robustness via Data Selection (NeurIPS 2024 · similarity 0.86)
- SEBRA : Debiasing through Self-Guided Bias Ranking (ICLR 2025 · similarity 0.84)
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · similarity 0.84)
- CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning (CVPR 2024 · similarity 0.84)
- LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing (AAAI 2025 · similarity 0.83)
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.
