Research map / Continual learning / Class-incremental
FG-OrIU: Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning
Qian Feng, JiaHang Tu, Mintong Kang, Hanbin Zhao, Chao Zhang, Hui Qian
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 |
|---|---|---|---|
| remove gradient orthogonal projection for remaining classes (add gop-r) (ablation) | FG-OrIU · CASIA-Face100 IU Task 1 | Acc_r: 71.45 -> 74.25 | Improved |
| add feature orthogonal projection for forgetting (fop-f) (ablation) | FG-OrIU · CASIA-Face100 IU Task 1 | Acc_f: 0.20 -> 0.00 | Improved |
| add gradient orthogonal projection for forgetting (gop-f) (ablation) | FG-OrIU · CASIA-Face100 IU Task 2 | Acc_f: 4.37 -> 0.00 | Improved |
| add feature orthogonal projection for remaining (fop-r) (ablation) | FG-OrIU · CASIA-Face100 IU Task 2 | Acc_r: 74.98 -> 75.50 | Mixed |
| add feature and gradient orthogonality to LoRA-based incremental unlearning (vs GS-LoRA) | GS-LoRA · CASIA-Face100 IU Task 4 | Acc_r: 75.36 -> 80.51 | Improved |
| add feature and gradient orthogonality (vs GS-LoRA) in static unlearning | GS-LoRA · CASIA-Face100 static MU (100-90) | Acc_r: 73.56 -> 77.15 | Improved |
| use FG-OrIU instead of DER++ on ImageNet100 Task 4 | DER++ · ImageNet100 IU Task 4 | H-Mean: 89.24 -> 89.15 | No clear effect |
| apply FG-OrIU instead of prior methods to hinder feature reconstruction | unlearned models · DIP reconstruction | SSIM: 0.75 -> 0.01 | Improved |
| freeze backbone and retrain head on prior unlearned models | existing IU methods · CASIA-Face100 recovery | forgetting-class accuracy recovered: 15%-70% of original performance recovered | Hurt |
Closest papers by meaning
- Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning (NeurIPS 2026 · similarity 0.87)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.87)
- Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · similarity 0.87)
- Machine Unlearning in Low-Dimensional Feature Subspace (NeurIPS 2026 · similarity 0.87)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.87)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.87)
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.
