Research map / Architectures and efficiency / Pruning and efficient tuning
On the Misalignment Between Data Learnability and Forgettability in Machine Unlearning
Zijie Pan, Zuobin Ying, Yajie Wang, Wanlei Zhou
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 |
|---|---|---|---|
| add Dual-Aware Training regularization (UGS variance and LFD penalty) to training | CNN · CIFAR-10 | LFD: 0.31 -> 0.15 (-52%) | Improved |
| add Dual-Aware Training to training | Transformer · IMDB | LFD: 0.28 -> 0.13 (-54%) | Improved |
| add Dual-Aware Training to training | CNN · MNIST | LFD: 0.21 -> 0.10 (-52%) | Improved |
| add Dual-Aware Training before unlearning | CNN · CIFAR-10 | unlearning latency (s): 3.84 -> 1.04 (3.7×) | Improved |
| add Dual-Aware Training before unlearning | Transformer · IMDB | unlearning latency (s): 0.90 -> 0.30 (3.0×) | Improved |
| add Dual-Aware Training to training | CNN · CIFAR-10 | accuracy (%): 93.1 -> 93.0 | No clear effect |
| add Dual-Aware Training to training | CNN · MNIST | accuracy (%): 99.1 -> 99.1 | No clear effect |
| add Dual-Aware Training instead of post-hoc SISA | CNN · CIFAR-10 | LFD: 0.18 -> 0.15 | Improved |
| use SISA unlearning on top of vanilla training (training overhead) | CNN · CIFAR-10 | training overhead (%): 45.0 vs 7.6 | Hurt |
| make regularization coefficients (lambda_ugs, lambda_lfd) very large (ablation) | DAT · CIFAR-10 | test accuracy | Hurt |
| make regularization coefficients too small (ablation) | DAT · CIFAR-10 | LFD and latency | Hurt |
| increase Hutch++ sketch dimension k beyond 4 (ablation) | DAT · CIFAR-10 | accuracy and LFD: k in {1, 2, 4, 8} | No clear effect |
Closest papers by meaning
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.91)
- Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning (NeurIPS 2026 · similarity 0.90)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.89)
- Fast Machine Unlearning without Retraining through Selective Synaptic Dampening (AAAI 2024 · similarity 0.89)
- RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 2025 · similarity 0.89)
- The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation (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.
