Research map / Optimization / Implicit bias and generalization
Sharpness-Aware Machine Unlearning
Haoran Tang, Rajiv Khanna
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 ASAM/SAM as the unlearning optimizer instead of SGD (ablation) | existing unlearning methods (NegGrad, RL, SalUn, SCRUB, L1-Sparse) · CIFAR-100 and ImageNet-1K | ToW: SCRUB 49.084 -> 73.766 (CIFAR-100, SGD pretrain) | Improved |
| pretrain with SAM/ASAM instead of SGD (ablation) | NegGrad unlearning · CIFAR-100 | ToW: 81.796 -> 84.789 (ASAM 0.1), 81.796 -> 72.78 (SAM 0.1) | Mixed |
| use Sharp MinMax (SAM on retain weights, sharpness maximization on forget weights) | unlearning with SAM 0.1 · CIFAR-100 | ToW: 83.514 -> 91.344 | Improved |
| use Sharp MinMax instead of NegGrad with ASAM 1.0 | ImageNet-1K SGD-pretrained · ImageNet-1K | ToW: 84.843 -> 87.899 | Improved |
| apply SAM/ASAM unlearning (vs SGD) to reduce MIA correctness on forget set (ablation) | L1-Sparse, SCRUB, RL, SalUn, NegGrad · CIFAR-100 | MIA correctness: SCRUB 35.544 -> 20.478 | Improved |
| use SAM instead of SGD for unlearning (feature entanglement) (ablation) | NegGrad · CIFAR-100 | E_Wp entanglement: 53.45 -> 52.39 | Improved |
| use SAM instead of SGD for unlearning (variance entanglement) (ablation) | NegGrad · CIFAR-100 | E_Var entanglement: 26.05 -> 26.58 | No clear effect |
| reduce NegGrad retain weight alpha (ablation) | SGD vs SAM/ASAM unlearning · CIFAR-100 | ToW | Hurt |
| use SAM vs SGD for unlearning: forget accuracy (ablation) | NegGrad · CIFAR-100 / ImageNet-1K | forget accuracy | Hurt |
| add SAM in unlearning on ViT-Small with AdamW (ablation) | NegGrad and Sharp MinMax · CIFAR-100 (ViT-Small) | ToW | Improved |
Closest papers by meaning
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.91)
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond (ICML 2025 · similarity 0.90)
- Forget Sharpness: Perturbed Forgetting of Model Biases Within SAM Dynamics (ICML 2024 · similarity 0.89)
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · similarity 0.89)
- SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion (NeurIPS 2026 · similarity 0.89)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.89)
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.
