Research map / Architectures and efficiency / Attention and transformers
FairViT: Fair Vision Transformer via Adaptive Masking
Bowei Tian, Ruijie Du, Yanning Shen
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 distance loss to cross-entropy (ablation) | FairViT · CelebA (Y: Attraction, S: Gender) | accuracy (%): 74.01 -> 77.01 | Improved |
| add static adaptive masks on top of distance loss (ablation) | FairViT · CelebA (Y: Attraction, S: Gender) | accuracy (%): 77.01 -> 79.96 | Improved |
| update masks and weights by gradient instead of keeping them static (ablation) | FairViT · CelebA (Y: Attraction, S: Gender) | accuracy (%): 79.96 -> 84.01 | Improved |
| update masks and weights by gradient instead of keeping them static (ablation) | FairViT · CelebA (Y: Attraction, S: Gender) | EO (e-2): 8.99 -> 1.15 | Improved |
| use static masks rather than none on fairness DP (ablation) | FairViT · CelebA (Y: Attraction, S: Gender) | DP (e-1): 3.166 -> 3.929 | Hurt |
| use FairViT instead of vanilla ViT | FairViT · CelebA (Y: Attraction, S: Gender) | EO (e-2): 14.43 -> 1.15 | Improved |
| use FairViT instead of FSCL+ for fairness | FairViT · CelebA (Y: Attraction, S: Gender) | EO (e-2): 0.79 -> 1.15 | Hurt |
| use FairViT instead of FSCL+ for accuracy | FairViT · CelebA (Y: Attraction, S: Gender) | accuracy (%): 77.26 -> 84.01 | Improved |
| increase distance loss weight alpha from 0.01 to 1 (ablation) | FairViT · CelebA (Y: Attraction, S: Gender) | accuracy (%): 82.51 -> 75.53 | Hurt |
| add small distance loss weight alpha=0.01 versus 0 (ablation) | FairViT · CelebA (Y: Attraction, S: Gender) | accuracy (%): 81.08 -> 82.51 | Improved |
| increase gamma beyond 0.5 (ablation) | FairViT · CelebA (Y: Attraction, S: Gender) | accuracy (%): 82.75 -> 81.67 | Hurt |
| use FairViT instead of FSCL+ for training time | FairViT · CelebA (Y: Attraction, S: Gender) | training time (min): 34.73 -> 5.09 | Improved |
Closest papers by meaning
- Fairness-aware Vision Transformer via Debiased Self-Attention (ECCV 2024 · similarity 0.92)
- Fair-VPT: Fair Visual Prompt Tuning for Image Classification (CVPR 2024 · similarity 0.87)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.85)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.85)
- Improving Interpretation Faithfulness for Vision Transformers (ICML 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.
