Research map / Label-efficient and robust learning / Recognition and long-tail
Discover and Mitigate Unknown Biases with Debiasing Alternate Networks
Zhiheng Li, Anthony Hoogs, Chenliang Xu
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 |
|---|---|---|---|
| add the unbalanced assignment (UA) penalty (ablation) | DebiAN · CelebA Blond Hair, worst group accuracy | accuracy: 38.5 -> 52.9 | Improved |
| add the UA penalty (ablation) | DebiAN · CelebA Blond Hair, average group accuracy | accuracy: 79.6 -> 84.0 | Improved |
| add the UA penalty (ablation) | DebiAN · Multi-Color MNIST unbiased accuracy | accuracy: 70.6 -> 72.0 | Improved |
| add the UA penalty (ablation) | DebiAN · CelebA gender classifier, Heavy Makeup worst group | accuracy: 46.9 -> 56.0 | Improved |
| add alternate training of discoverer and classifier (ablation) | DebiAN · Multi-Color MNIST unbiased accuracy | accuracy: 70.7 -> 72.0 | Improved |
| add alternate training (ablation) | DebiAN · Multi-Color MNIST, bias-conflicting on both colors | accuracy: 11.5 -> 16.0 | Improved |
| use reweighted cross-entropy instead of minmax game on the classifier (ablation) | DebiAN · Multi-Color MNIST unbiased accuracy | accuracy: 66.3 -> 72.0 | Improved |
| predict bias groups per class instead of globally (ablation) | discoverer · Multi-Color MNIST, bias-conflicting on both colors | accuracy: 13.5 -> 16.0 | Improved |
| predict bias groups per class instead of globally (ablation) | discoverer · bFFHQ | accuracy: 62.80 vs 62.87 | No clear effect |
| use DebiAN instead of LfF in the multi-bias setting | debiasing method · Multi-Color MNIST unbiased accuracy | accuracy: 52.0 -> 72.0 | Improved |
| use DebiAN instead of vanilla training with batch size 512 | classifier · Multi-Color MNIST unbiased accuracy | accuracy: 56.1 -> 70.1 | Improved |
| use DebiAN instead of vanilla training on BAR | classifier · Biased Action Recognition | accuracy: 51.85 -> 69.88 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BiaSwap — BiaSwap: Removing Dataset Bias With Bias-Tailored Swapping Augmentation (ICCV 2021) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Fast Model DeBias with Machine Unlearning (NeurIPS 2023 · similarity 0.89)
- Learning Debiased Representation via Disentangled Feature Augmentation (NeurIPS 2021 · similarity 0.88)
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · similarity 0.87)
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.87)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.86)
- Discover and Mitigate Multiple Biased Subgroups in Image Classifiers (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.
