Research map / Optimization / Optimizers
Learning to Unlearn while Retaining: Combating Gradient Conflicts in Machine Unlearning
Gaurav Patel, Qiang Qiu
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 the inner retain-gradient step (LUR objective) instead of plainly summing retain and forget losses (ablation) | LUR · CIFAR-10 random forgetting 10% | Avg. Gap to Retrain: 3.15 -> 0.89 | Improved |
| add the inner retain-gradient step (LUR objective) instead of plainly summing retain and forget losses (ablation) | LUR · CIFAR-10 random forgetting 50% | Avg. Gap to Retrain: 3.51 -> 2.62 | Improved |
| add the inner retain-gradient step (LUR objective) instead of plainly summing retain and forget losses (ablation) | LUR · CIFAR-100 random forgetting 10% | Avg. Gap to Retrain: 4.28 -> 3.96 | Improved |
| add the inner retain-gradient step (LUR objective) instead of plainly summing retain and forget losses (ablation) | LUR · CIFAR-100 random forgetting 50% | Avg. Gap to Retrain: 15.64 -> 8.07 | Improved |
| use LUR instead of SHs, explicit gradient projection, for identity forgetting | ResNet-34 · Celeb-HQ-FIR identity forgetting 10% | Avg. Gap to Retrain: 2.82 -> 0.10 | Improved |
| use LUR instead of SalUn for 50% random forgetting on CIFAR-10 | ResNet-18 · CIFAR-10 random forgetting 50% | Avg. Gap to Retrain: 2.18 -> 2.62 | Hurt |
| use LUR instead of SalUn for CIFAR-100 forgetting | ResNet-18 · CIFAR-100 random forgetting 10% | Avg. Gap to Retrain: 10.45 -> 3.96 | Improved |
| use LUR instead of SA for DDPM class unlearning | DDPM · CIFAR-10 class-wise forgetting (DDPM) | UA: 85.80 -> 100.00 | Improved |
| use LUR instead of SA for DDPM class unlearning | DDPM · CIFAR-10 class-wise forgetting (DDPM) | FID: 9.08 -> 9.76 | Hurt |
| use LUR instead of SalUn for Stable Diffusion class unlearning | Stable Diffusion · Imagenette class-wise forgetting (average) | FID: 1.22 -> 0.98 | Improved |
| vary the inner learning step alpha; too large or too small (ablation) | LUR · Classification unlearning | Avg. Gap to Retrain: optimal alpha = 0.01 | Mixed |
Closest papers by meaning
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.91)
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.91)
- RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 2025 · similarity 0.90)
- What makes unlearning hard and what to do about it (NeurIPS 2024 · similarity 0.90)
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.90)
- Adversarial Mixup Unlearning (ICLR 2025 · similarity 0.90)
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.
