Research map / Continual learning / Class-incremental
POUR: A Provably Optimal Method for Unlearning Representation via Neural Collapse
Anjie Le, Can Peng, Yuyuan Liu, J. Alison Noble
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 |
|---|---|---|---|
| distill feature extractor on forget set toward projected features (POUR-D) instead of post-hoc projection only (POUR-P) (ablation) | POUR · CIFAR-10 (ResNet18) | Acc_r: 94.97 -> 92.86 | Hurt |
| use POUR-D to modify encoder representations (POUR-P leaves encoder unchanged) (ablation) | POUR · CIFAR-10 (ResNet18) | rMIA: 56.67 -> 51.80 | Improved |
| POUR-D versus best forget-set-only baselines on representation metrics | POUR-D · CIFAR-10 (ResNet18) | RUS(o): 0.80 -> 0.85 | Improved |
| POUR-D versus DELETE on retain accuracy | POUR-D · CIFAR-10 (ResNet18) | Acc_r: 88.73 -> 92.86 | Improved |
| POUR-P versus Gradient Ascent on CIFAR-100 | POUR-P · CIFAR-100 (ResNet18) | Acc_r: 50.46 -> 77.65 | Improved |
| use POUR-D instead of POUR-P on CIFAR-100 (ablation) | POUR · CIFAR-100 (ResNet18) | Acc_r: 77.65 -> 73.44 | Hurt |
| weak class separation (CIFAR-100 entangled classes) makes forget-set-only unlearning harder | forget-set-only unlearning · CIFAR-100 (ResNet18) | RUS(r): 0.65 (POUR-D) vs 0.84 on CIFAR-10 RUS(r) not reported; CIFAR-10 POUR-D 0.47 | Hurt |
| apply POUR to pretrained ViT with domain shift | POUR-P · PathMNIST internal vs external | Acc_r: 87.14 internal, 87.44 external | Improved |
| Random Label and Gradient Ascent on PathMNIST | baselines · PathMNIST | Acc_r internal vs external: 81.43 -> 76.72 (Gradient Ascent) | Hurt |
| POUR-P on CLIP text-embedding removal for ImageNet classes | POUR-P · ImageNet (CLIP-L/14) | Acc_f: 94.00 -> 14.00 (Goldfish) | Improved |
| POUR on semantic segmentation forgetting | DeepLabV3+ · VOC 2012 | IoU_f and IoU_r: Dog IoU_f 90.98 -> 0.00; IoU_r 77.81 -> 71.53 | Mixed |
Closest papers by meaning
- Label-Agnostic Forgetting: A Supervision-Free Unlearning in Deep Models (ICLR 2024 · similarity 0.89)
- Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · similarity 0.89)
- Targeted Forgetting of Image Subgroups in CLIP Models (CVPR 2025 · similarity 0.88)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.88)
- The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation (NeurIPS 2026 · similarity 0.88)
- Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement (NeurIPS 2024 · similarity 0.88)
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.
