Research map / Learning theory and algorithms / Prediction and decision losses
What makes unlearning hard and what to do about it
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies embedding-space entanglement and memorization as factors that make forget sets hard to unlearn, and proposes RUM, which refines forget sets by memorization and unlearns the subsets sequentially with existing algorithms.
Problem
Machine unlearning lacks understanding of which interpretable characteristics of forget sets make removing their influence hard, and how state-of-the-art approximate unlearning algorithms behave on such forget sets.
Approach
Identifies two interpretable difficulty factors (retain/forget embedding-space entanglement and memorization score), shows that algorithms diverge across these factors, and proposes RUM: partition the forget set by memorization and apply existing unlearning algorithms sequentially to each homogeneous subset, optionally choosing a different algorithm per subset.
Key results
- On CIFAR-10 with ResNet-18, RUM (nothing -> Fine-tune -> SalUn) reaches ToW 0.965±0.014 with MIA gap 0.034, versus 0.731 ToW for vanilla SalUn.
- Applying NegGrad+ as RUM^F on CIFAR-100 raises ToW from 0.861 (vanilla) to 0.921, and MIA gap falls from 0.159 to 0.059.
- Relabelling-based methods (SalUn, Random-label) degrade as embedding entanglement increases and show a reverse trend with memorization compared to other methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Refine the forget set by memorization and apply Fine-tune sequentially to homogeneous subsets instead of once to the full set (ablation) | Fine-tune on CIFAR-10 · CIFAR-10 | ToW: 0.849±0.030 -> 0.937±0.052 | Helped |
| Apply Fine-tune sequentially to random subsets instead of once to the full set (ablation) | Fine-tune on CIFAR-10 · CIFAR-10 | ToW: 0.849±0.030 -> 0.712±0.040 | Hurt |
| Refine the forget set by memorization and apply L1-sparse sequentially to homogeneous subsets instead of once to the full set (ablation) | L1-sparse on CIFAR-10 · CIFAR-10 | ToW: 0.794±0.035 -> 0.900±0.020 | Helped |
| Apply L1-sparse sequentially to random subsets instead of once to the full set (ablation) | L1-sparse on CIFAR-10 · CIFAR-10 | ToW: 0.794±0.035 -> 0.716±0.023 | Hurt |
| Refine the forget set by memorization and apply NegGrad+ sequentially to homogeneous subsets instead of once to the full set (ablation) | NegGrad+ on CIFAR-10 · CIFAR-10 | ToW: 0.802±0.028 -> 0.879±0.068 | Helped |
| Apply NegGrad+ sequentially to random subsets instead of once to the full set (ablation) | NegGrad+ on CIFAR-10 · CIFAR-10 | ToW: 0.802±0.028 -> 0.632±0.022 | Hurt |
| Refine the forget set by memorization and apply SalUn sequentially to homogeneous subsets instead of once to the full set (ablation) | SalUn on CIFAR-10 · CIFAR-10 | ToW: 0.731±0.070 -> 0.887±0.069 | Helped |
| Select an unlearning algorithm per memorization subset, using nothing, Fine-tune, then SalUn (ablation) | RUM on CIFAR-10 · CIFAR-10 | ToW: 0.887±0.069 -> 0.965±0.014 | Helped |
| Refine the forget set by memorization and apply Fine-tune sequentially to homogeneous subsets instead of once to the full set (ablation) | Fine-tune on CIFAR-100 · CIFAR-100 | ToW: 0.734±0.025 -> 0.784±0.040 | Helped |
| Apply Fine-tune sequentially to random subsets instead of once to the full set (ablation) | Fine-tune on CIFAR-100 · CIFAR-100 | ToW: 0.734±0.025 -> 0.589±0.036 | Hurt |
| Refine the forget set by memorization and apply L1-sparse sequentially to homogeneous subsets instead of once to the full set (ablation) | L1-sparse on CIFAR-100 · CIFAR-100 | ToW: 0.824±0.011 -> 0.883±0.046 | Helped |
| Refine the forget set by memorization and apply NegGrad+ sequentially to homogeneous subsets (ablation) | NegGrad+ on CIFAR-100 · CIFAR-100 | ToW: 0.861±0.069 -> 0.921±0.034 | Helped |
Surprising findings
- Relabelling-based methods like SalUn perform worst on low-memorization forget sets, where doing nothing is already near-optimal.
- Sequential 'shuffle' unlearning usually performs worse than vanilla one-shot unlearning, so a poor refinement can increase difficulty.
Limitations
- Experiments limited to image classification on CIFAR and Tiny-ImageNet with small ResNet/VGG models; no LLM experiments.
- Only memorization-based refinement and sequential meta-algorithm instantiations were evaluated; embedding-entanglement-based refinement left for future work.
- Hyperparameters were tuned per subset to match retrain accuracy, which may not be practical in deployment.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet
Metrics: ToW MIA gap forget accuracy retain accuracy test accuracy ES per-example disagreement
Baselines: Fine-tune NegGrad NegGrad+ L1-sparse SalUn Random-label Influence unlearning Retrain
Builds on: SalUn SCRUB NegGrad+ Memorization score (Feldman & Zhang) Influence unlearning
Closest papers by meaning
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.93)
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.90)
- System-Aware Unlearning Algorithms: Use Lesser, Forget Faster (ICML 2025 · similarity 0.89)
- Towards Source-Free Machine Unlearning (CVPR 2025 · similarity 0.89)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.89)
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · 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.