Research map / Learning theory and algorithms / Prediction and decision losses
A Near-Optimal Algorithm for Debiasing Trained Machine Learning Models
Ibrahim Alabdulmohsin, Mario Lucic
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 randomized threshold post-processing to enforce statistical parity | Random forests, k-NN, MLP, and logistic regression · Adult test | bias: Original: .38, .24, .29, .39; RTO: .01, .02, .01, .01 | Improved |
| Apply randomized threshold post-processing to enforce statistical parity | Random forests, k-NN, MLP, and logistic regression · DCCC test | bias: Original: .07, .10, .13, .12; RTO: .01, .01, .01, .01 | Improved |
| Apply randomized threshold post-processing to enforce statistical parity | ResNet50 and MobileNet · CelebA test | bias: Original: .43, .40, .35, .38; RTO: .01, 0.02, .01, .002 | Improved |
| Use randomized rather than deterministic thresholding | RTO versus ROC · CelebA test | bias: ResNet50 scratch: .08 -> .01; ResNet50 ImageNet: .15 -> 0.02; MobileNet ImageNet: .10 -> .002 | Improved |
| Post-process the model with RTO | Random forests, k-NN, MLP, and logistic regression · DCCC test | accuracy: 81.2 -> 81.8; 79.6 -> 80.4; 80.5 -> 81.3; 80.6 -> 81.7 | Improved |
| Post-process the model with RTO | Random forests, k-NN, MLP, and logistic regression · Adult test | accuracy: 85.7 -> 84.4; 86.8 -> 81.3; 85.5 -> 83.5; 84.9 -> 83.0 | Hurt |
| Post-process the model with RTO | ResNet50 and MobileNet · CelebA test | accuracy: 77.8 -> 71.3; 79.7 -> 71.7; 76.9 -> 71.8; 79.3 -> 72.8 | Hurt |
| Increase pretrained model size | Debiased pretrained DNNs · CelebA and COCO tasks | normalized error rate | Improved |
| Increase upstream pretraining dataset size | BiT models and ResNet50 · CelebA and COCO tasks | test error | Improved |
| Pretrain BiT models on JFT-300M rather than the other evaluated upstream datasets | BiT models · CelebA and COCO tasks | average ranking by test error | Improved |
Closest papers by meaning
- Debiasing surgeon: fantastic weights and how to find them (ECCV 2024 · similarity 0.87)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.86)
- Limits to scalable evaluation at the frontier: LLM as judge won’t beat twice the data (ICLR 2025 · similarity 0.86)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.85)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.85)
- SEBRA : Debiasing through Self-Guided Bias Ranking (ICLR 2025 · 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.
