Research map / Label-efficient and robust learning / Recognition and long-tail
Fast Model DeBias with Machine Unlearning
Ruizhe Chen, Jianfei Yang, Huimin Xiong, Jianhong Bai, Tianxiang Hu, Jin Hao, YANG FENG, Joey Tianyi Zhou, Jian Wu, Zuozhu Liu
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 FMD unlearning to the biased model | Colored MNIST, bias ratio 0.995 · Colored MNIST (bias ratio 0.995) | bias: 0.5863 -> 0.3027 | Improved |
| Apply FMD unlearning to the biased model | Colored MNIST, bias ratio 0.99 · Colored MNIST (bias ratio 0.99) | bias: 0.4931 -> 0.2042 | Improved |
| Apply FMD unlearning to the biased model | Colored MNIST, bias ratio 0.95 · Colored MNIST (bias ratio 0.95) | bias: 0.2589 -> 0.1189 | Improved |
| Apply FMD unlearning | Adult Income, Gender · Adult Income (Gender) | bias: 0.0195 -> 0.0005 | Improved |
| Apply FMD unlearning | Adult Income, Race · Adult Income (Race) | bias: 0.0089 -> 0.0013 | Improved |
| Apply FMD unlearning | CelebA (Blonde) · CelebA (Blonde) | worst-group accuracy: 47.36 -> 87.15 | Improved |
| Apply FMD unlearning | CelebA (Attractive) · CelebA (Attractive) | worst-group accuracy: 40.59 -> 68.16 | Improved |
| Use counterfactual bias rather than demographic-parity bias to guide unlearning (ablation) | Colored MNIST, bias severity 0.99 · Colored MNIST (bias severity 0.99) | accuracy: 71.52 -> 87.90 | Improved |
| Use counterfactual bias rather than equal-opportunity bias to guide unlearning (ablation) | Colored MNIST, bias severity 0.99 · Colored MNIST (bias severity 0.99) | accuracy: 71.12 -> 87.90 | Improved |
| Replace harmful training examples with bias-conflicting examples (ablation) | FMD unlearning · Colored MNIST (bias severity 0.99) | bias: 0.1182 -> 0.1023 | Improved |
| Use external counterfactual pairs instead of training examples (ablation) | FMD unlearning · Colored MNIST (bias severity 0.99) | debiasing time: 39.63 -> 0.059 | Improved |
| Update 2 rather than 1 MLP layer (ablation) | FMD · Colored MNIST (bias ratio 0.99) | accuracy: 71.19 -> 74.18 | Improved |
Closest papers by meaning
- A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective (NeurIPS 2024 · similarity 0.88)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.88)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.87)
- MABR: Multilayer Adversarial Bias Removal Without Prior Bias Knowledge (AAAI 2025 · similarity 0.87)
- SEBRA : Debiasing through Self-Guided Bias Ranking (ICLR 2025 · similarity 0.87)
- An Effective Theory of Bias Amplification (ICLR 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.
