Research map / Learning theory and algorithms / Prediction and decision losses
Algorithms that Approximate Data Removal: New Results and Limitations
Vinith Menon Suriyakumar, Ashia Camage Wilson
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 IJ instead of retraining from scratch | unlearning with l2 logistic regression · MNIST 3 vs 8 | runtime: 2611x | Improved |
| use IJ instead of Algorithm TA (Hessian inverted once, not per request) | unlearning with l2 logistic regression · MNIST 3 vs 8 | runtime: 2045x | Improved |
| use IJ instead of Algorithm TA | unlearning, test accuracy · MNIST 3 vs 8 | test accuracy | No clear effect |
| adaptive instead of random delete requests | IJ · MNIST 3 vs 8 | test accuracy | No clear effect |
| use IJ (proximal Newton) instead of retraining with l1 regularizer | unlearning with l1 logistic regression · Warfarin dosing | runtime: 30x | Improved |
| use IJ instead of retraining, l1 regularizer | unlearning with l1 logistic regression · Warfarin dosing | test accuracy | Mixed |
| use IJ instead of retraining on non-convex DP feature extractor | unlearning · SVHN 3 vs 8 | runtime: 15036x | Improved |
| use IJ instead of TA on non-convex DP feature extractor | unlearning · SVHN 3 vs 8 | runtime: 17x | Improved |
| use IJ on non-convex setting as deletions increase | unlearning · SVHN 3 vs 8 | test accuracy | No clear effect |
| select lambda by cross-validation before unlearning | approximate unlearning guarantees · theory counterexample | unlearning guarantee | Hurt |
Closest papers by meaning
- Hessian-Free Online Certified Unlearning (ICLR 2025 · similarity 0.90)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.89)
- The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning (ICLR 2025 · similarity 0.89)
- Rewind-to-Delete: Certified Machine Unlearning for Nonconvex Functions (NeurIPS 2025 · similarity 0.88)
- Towards Scalable Exact Machine Unlearning Using Parameter-Efficient Fine-Tuning (ICLR 2025 · similarity 0.88)
- Machine Unlearning under Overparameterization (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.
