Research map / Learning theory and algorithms / Fairness
Leveling Down in Computer Vision: Pareto Inefficiencies in Fair Deep Classifiers
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 g-SMOTE augmentation with adaptive sampling | CelebA classifiers · CelebA, gender-independent labels | Min. grp. acc.: 90.14 -> 90.36 | Helped |
| Apply common fairness methods instead of the corresponding baseline | CelebA classifiers · CelebA, gender-independent labels | Min. grp. acc.: 89.52 -> 89.06; 89.39; 88.93; 88.08 | Hurt |
| Apply GAN debiasing or DEO regularization instead of the single-task baseline | CelebA classifiers · CelebA, gender-independent labels | Min. grp. acc.: 90.14 -> 89.85; 87.86 | Hurt |
| Apply FairMixup | CelebA classifiers · CelebA, gender-independent labels | Min. grp. acc.: 91.44 -> 86.36 | Hurt |
| Add g-SMOTE augmentation with adaptive sampling | CelebA classifiers · CelebA, 2 protected groups, 10k train images | Min. grp. acc.: 85.68 -> 85.93 | Helped |
| Remove adaptive sampling while retaining g-SMOTE (ablation) | Adaptive g-SMOTE · CelebA, 2 protected groups, 10k train images | Min. grp. acc.: 85.93 -> 85.91 | No clear effect |
| Add g-SMOTE augmentation with adaptive sampling | CelebA classifiers · CelebA, 4 protected groups, 10k train images | Min. grp. acc.: 62.16 -> 64.36 | Helped |
| Remove adaptive sampling while retaining g-SMOTE (ablation) | Adaptive g-SMOTE · CelebA, 4 protected groups, 10k train images | Min. grp. acc.: 64.36 -> 63.54 | Hurt |
| Add g-SMOTE augmentation with adaptive sampling | CelebA classifiers · CelebA, 4 protected groups, 160k train images | Min. grp. acc.: 61.74 -> 68.06 | Helped |
| Remove adaptive sampling while retaining g-SMOTE (ablation) | Adaptive g-SMOTE · CelebA, 4 protected groups, 160k train images | Min. grp. acc.: 68.06 -> 61.84 | Hurt |
| Add g-SMOTE augmentation (ablation) | CelebA classifiers · CelebA, gender-independent labels, no augmentation | Min. group accuracy: 89.15 -> 89.63 | Helped |
| Add g-SMOTE to RandAugment (ablation) | CelebA classifiers · CelebA, gender-independent labels, RandAugment | Min. group accuracy: 90.17 -> 90.33 | Helped |
Closest papers by meaning
- Classes Are Not Equal: An Empirical Study on Image Recognition Fairness (CVPR 2024 · similarity 0.89)
- Some Optimizers are More Equal: Understanding the Role of Optimizers in Group Fairness (NeurIPS 2025 · similarity 0.88)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.88)
- The Disparate Benefits of Deep Ensembles (ICML 2025 · similarity 0.88)
- On the Fairness ROAD: Robust Optimization for Adversarial Debiasing (ICLR 2024 · similarity 0.87)
- Efficient Fairness-Performance Pareto Front Computation (NeurIPS 2025 · similarity 0.87)
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.