Research map / Continual learning / Continual fine-tuning of LLMs
Forgetting Knowledge Localization and Isolation for Continual Forgetting of Pre-trained Vision Models
Zhiwen Yang, Jiehua Zhang, Chenggang Yan, Yuhan Gao, Zongpeng Li, Xichun Sheng, Liang Li
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 FKPI parameter isolation alone (ablation) | continual forgetting baseline · ImageNet100 | H: 84.79 -> 86.16 | Improved |
| add FKLL layer localization alone (ablation) | continual forgetting baseline · ImageNet100 | H: 84.79 -> 83.06 | Hurt |
| add all four components (FKLL, FKPI, LFKIL, LAlign) (ablation) | continual forgetting baseline · ImageNet100 | H: 84.79 -> 87.01 | Improved |
| add FKLL on top of FKPI (ablation) | FKPI · ImageNet100 | H: 86.16 -> 85.23 | Hurt |
| adaptive layer localization instead of random layers (ablation) | FKLL · ImageNet100 | H: 81.37 -> 87.01 | Improved |
| adaptive per-task layer selection instead of fixed top-12 layers (ablation) | FKLL · ImageNet100 | H: 85.01 -> 87.01 | Improved |
| SVD-based sensitivity initialization of LoRA instead of zero init (ablation) | FKPI · ImageNet100 | H: 82.98 -> 87.01 | Improved |
| SVD-based initialization instead of standard LoRA initialization (ablation) | FKPI · ImageNet100 | H: 72.98 -> 87.01 | Improved |
| increase LoRA rank (ablation) | FKPI · ImageNet100 | H | Mixed |
| use our method instead of GS-LoRA | continual forgetting · CASIA-Face100 (avg of 4 tasks) | H: 70.0 -> 74.7 | Improved |
| use our method instead of GS-LoRA | continual forgetting · ImageNet100 (avg of 4 tasks) | H: 85.9 -> 87.0 | Improved |
Closest papers by meaning
- Don't Forget Why You Started: Tackling Dual Forgetting in Vision-Language Continual Learning (ICML 2026 · similarity 0.90)
- Practical Continual Forgetting for Pre-Trained Vision Models (TPAMI 2026 · similarity 0.89)
- KeepLoRA: Continual Learning with Residual Gradient Adaptation (ICLR 2026 · similarity 0.89)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.87)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.87)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · 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.
