Research map / Label-efficient and robust learning / Recognition and long-tail
Fair Generative Modeling via Weak Supervision
Read the paperCheck 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 |
|---|---|---|---|
| Use classifier-estimated density ratios to reweight training examples | BigGAN fair generation · CelebA, single bias=0.9 | fairness discrepancy: 49.3% over equi-weight; 25.0% over conditional | Helped |
| Use classifier-estimated density ratios to reweight training examples | BigGAN fair generation · CelebA, single bias=0.8 | fairness discrepancy: 23.9% over equi-weight; 12.2% over conditional | Helped |
| Use classifier-estimated density ratios to reweight training examples | BigGAN fair generation · CelebA, multi-attribute bias | fairness discrepancy: 32.5% over equi-weight; 4.4% over conditional | Helped |
| Use classifier-estimated density ratios to reweight training examples | GAN fair generation · CelebA | fairness discrepancy: up to 34.6% | Helped |
| Importance-reweight the biased training data | GAN fair generation · Shapes3D, floor-color bias=0.9 | fairness discrepancy: 0.468 -> 0.002 | Helped |
| Use classifier-estimated density ratios to weight generated-data training | AC-GAN downstream classifier · CelebA downstream classification, bias=0.9, perc=1.0 | demographic parity distance: 0.032 -> 0.002 vs equi-weight | Helped |
| Use classifier-estimated density ratios to weight generated-data training | AC-GAN downstream classifier · CelebA downstream classification, bias=0.9, perc=1.0 | accuracy: 79% -> 75% vs equi-weight | Hurt |
| Use classifier-estimated density ratios to weight generated-data training | AC-GAN downstream classifier · CelebA downstream classification, bias=0.9, perc=1.0 | NLL: 0.7902 -> 0.7564 vs equi-weight | Helped |
| Condition the generator on whether an example came from the reference or biased dataset instead of importance-reweighting | BigGAN fair generation · CelebA single- and multi-attribute splits | fairness discrepancy: 25.0% lower with imp-weight at bias=0.9; 12.2% at bias=0.8; 4.4% in multi-attribute setting | Hurt |
| Train only on the reference dataset rather than combining it with biased data | BigGAN fair generation · CelebA | recognizable sample generation | Hurt |
| Flatten density ratios by using a power other than the standard exponent (ablation) | BigGAN fair generation · CelebA | fairness discrepancy | Mixed |
| Remove importance reweighting (α=0, equivalent to equi-weight) (ablation) | BigGAN fair generation · CelebA | fairness discrepancy: imp-weight outperforms equi-weight by 49.3% at bias=0.9, 23.9% at bias=0.8, and 32.5% in multi-attribute setting | Hurt |
Closest papers by meaning
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.88)
- Low-Biased General Annotated Dataset Generation (CVPR 2025 · similarity 0.87)
- Fusing Conditional Submodular GAN and Programmatic Weak Supervision (AAAI 2024 · similarity 0.87)
- Would Deep Generative Models Amplify Bias in Future Models? (CVPR 2024 · similarity 0.87)
- Balancing Act: Distribution-Guided Debiasing in Diffusion Models (CVPR 2024 · similarity 0.86)
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 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.