Research map / Label-efficient and robust learning / Recognition and long-tail
A Conservative Approach for Unbiased Learning on Unknown Biases
Myeongho Jeon, Daekyung Kim, Woochul Lee, Myungjoo Kang, Joonseok Lee
Check 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 |
|---|---|---|---|
| Use UBNet instead of the base classifier | VGG11 on CelebA-HQ · val-EB1 | accuracy: 99.38(±0.31) -> 99.18(±0.18) | No clear effect |
| Use UBNet instead of the base classifier | VGG11 on CelebA-HQ · val-EB2 | accuracy: 51.22(±1.73) -> 58.22(±0.64) | Improved |
| Use UBNet instead of the base classifier | VGG11 on CelebA-HQ · test set | accuracy: 75.30(±0.93) -> 78.70(±0.24) | Improved |
| Use UBNet instead of the strongest listed baseline | ResNet18 on UTKFace skin tone prediction · trained on UB1, evaluated on UB2 | accuracy: 79.35(±0.17) -> 83.67(±1.05) | Improved |
| Use UBNet instead of the strongest listed baseline | ResNet18 on UTKFace skin tone prediction · trained on UB1, test set | accuracy: 85.07(±0.46) -> 87.25(±0.82) | Improved |
| Use UBNet instead of the strongest listed baseline | ResNet18 on UTKFace skin tone prediction · trained on UB2, evaluated on UB1 | accuracy: 80.82(±0.15) -> 84.29(±1.24) | Improved |
| Use UBNet instead of the strongest listed baseline | ResNet18 on UTKFace skin tone prediction · trained on UB2, test set | accuracy: 85.88(±0.27) -> 87.94(±0.80) | Improved |
| Use UBNet instead of the strongest listed baseline | ResNet18 on UTKFace gender prediction · trained on UB1, evaluated on UB2 | accuracy: 82.02(±0.84) -> 82.07(±1.55) | No clear effect |
| Use UBNet instead of the strongest listed baseline | ResNet18 on UTKFace gender prediction · trained on UB1, test set | accuracy: 83.67(±0.73) -> 87.75(±0.58) | Improved |
| Add hierarchical features, group convolution, and orthogonal regularization (ablation) | VGG11-based classifier on CelebA-HQ · test set | accuracy: 76.11 -> 78.70 | Improved |
| Add group convolution to the hierarchical-feature model (ablation) | VGG11-based classifier on CelebA-HQ · test set | accuracy: 77.72 -> 77.93 | Improved |
| Add orthogonal regularization to hierarchical features and group convolution (ablation) | VGG11-based classifier on CelebA-HQ · test set | accuracy: 77.93 -> 78.70 | Improved |
Closest papers by meaning
- Unbiased Supervised Contrastive Learning (ICLR 2023 · similarity 0.87)
- BiaSwap: Removing Dataset Bias With Bias-Tailored Swapping Augmentation (ICCV 2021 · similarity 0.87)
- Discover and Mitigate Unknown Biases with Debiasing Alternate Networks (ECCV 2022 · similarity 0.87)
- Learning Debiased Representation via Disentangled Feature Augmentation (NeurIPS 2021 · similarity 0.87)
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · similarity 0.85)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.84)
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.
