Research map / Architectures and efficiency / Attention and transformers
Fairness-aware Vision Transformer via Debiased Self-Attention
Yao Qiang, Chengyin Li, Prashant Khanduri, Dongxiao Zhu
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 |
|---|---|---|---|
| apply DSA instead of vanilla ViT training | ViT-S/16 · CelebA (Hair Color, Gender) | DBA: 0.0422 -> 0.0031 | Improved |
| apply DSA instead of vanilla ViT training | ViT-S/16 · Waterbirds | Accuracy (%): 62.36 -> 69.58 | Improved |
| use CNN-designed debiasing (MMD, MFD, DANN) on ViT | ViT-S/16 · Waterbirds | EO: 0.0209 -> 0.0617 (MMD) | Hurt |
| add CE loss on original images to DSA objective (ablation) | DSA · CelebA | Accuracy (%): 88.32 -> 90.95 | Improved |
| add CE loss on adversarial examples to DSA objective (ablation) | DSA · CelebA | Accuracy (%): 88.54 -> 90.95 | Improved |
| add attention weights alignment regularizer (ablation) | DSA · CelebA | DP: 0.2865 -> 0.2337 | Improved |
| add attention weights alignment regularizer (ablation) | DSA · Waterbirds | EO: 0.0285 -> 0.0185 | Improved |
| perturb 5 patches instead of 3 (ablation) | DSA · CelebA | Accuracy (%): 90.95 -> 88.55 | Hurt |
| perturb only 1 patch instead of 3 (ablation) | DSA · CelebA | DP: 0.2337 -> 0.3075 | Hurt |
| use AT discrepancy instead of MSE for attention alignment (ablation) | DSA · CelebA | EO: 0.2706 -> 0.2558 | Improved |
| use AT instead of KL-Div for attention alignment (ablation) | DSA · CelebA | Balanced Accuracy: 83.26 -> 82.92 | Mixed |
| use attention masking from bias-only model instead of adversarial attack (AM vs DSA) (ablation) | DSA · CelebA | DBA: 0.0256 -> 0.0031 | Improved |
Closest papers by meaning
- FairViT: Fair Vision Transformer via Adaptive Masking (ECCV 2024 · similarity 0.92)
- Debiasing Attention Mechanism in Transformer without Demographics (ICLR 2024 · similarity 0.87)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.86)
- Fair-VPT: Fair Visual Prompt Tuning for Image Classification (CVPR 2024 · similarity 0.86)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.85)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · 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.
