Research map / Learning theory and algorithms / Prediction and decision losses
Selective Data Removal for Distributional Machine Unlearning
Youssef Allouah, Rachid Guerraoui, Sanmi Koyejo
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 |
|---|---|---|---|
| use selective distance-based removal instead of random removal (ablation) | distributional unlearning · low-divergence Gaussians | deletion budget (%): 65 -> 18 | Improved |
| use selective removal instead of random removal (ablation) | distributional unlearning · high-divergence Gaussians | deletion budget (%): 65 -> 50 | Improved |
| use selective removal (LR-COS) instead of random removal (ablation) | distributional unlearning · Jigsaw toxic comments | deletion budget (%) to halve recall: 100 -> 85 | Improved |
| use selective removal instead of random removal (ablation) | distributional unlearning · SMS Spam | deletion budget (%): 90 -> 75 | Improved |
| use selective removal (LR-MAHA) instead of random removal (ablation) | distributional unlearning · CIFAR-10 cat class | deletion budget (%): 80 -> 50 | Improved |
| use selective removal instead of random selection before retraining (ablation) | retraining · CIFAR-10 cat unlearning | deletion budget (%): 80 -> 60 | Improved |
| use selective removal instead of random selection before finetuning (ablation) | finetuning · CIFAR-10 cat unlearning | deletion budget (%): 98 -> 88 | Improved |
| use selective removal instead of random selection before SalUn (ablation) | SalUn · CIFAR-10 cat unlearning | deletion budget (%): 68 -> 55 | Improved |
| use selective removal instead of coreset selection before SalUn (ablation) | SalUn · CIFAR-10 cat unlearning | deletion budget (%): 75 -> 55 | Improved |
| use selective removal instead of coreset selection with influence functions (ablation) | influence-function unlearning · low-divergence Gaussians | deletion budget (%): 93 -> 18 | Improved |
| use centroid coreset instead of random selection with retraining (ablation) | retraining · CIFAR-10 cat unlearning | deletion budget (%): 80 -> 75 | Improved |
| use centroid coreset instead of random selection with influence functions (ablation) | influence-function unlearning · low-divergence Gaussians | deletion budget (%): 91 -> 93 | Hurt |
Closest papers by meaning
- Towards Source-Free Machine Unlearning (CVPR 2025 · similarity 0.90)
- The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning (ICLR 2025 · similarity 0.90)
- Data Unlearning in Diffusion Models (ICLR 2025 · similarity 0.89)
- SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion (NeurIPS 2026 · similarity 0.88)
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.88)
- Exact Unlearning via Quantized Sufficient Statistics (NeurIPS 2026 · similarity 0.88)
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.
