Research map / Continual learning / Continual fine-tuning of LLMs
Practical Continual Forgetting for Pre-Trained Vision Models
Hongbo Zhao, Fei Zhu, Bolin Ni, Feng Zhu, Gaofeng Meng, Zhaoxiang Zhang
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 prototype regularization (GS-LoRA to GS-LoRA++) | GS-LoRA · CASIA-Face100 4-shot single-step 100-10 | H-Mean: 63.61 -> 69.47 | Improved |
| add prototype regularization (GS-LoRA to GS-LoRA++) | GS-LoRA · CASIA-Face100 4-shot single-step 100-90 | H-Mean: 26.82 -> 60.86 | Improved |
| add prototype regularization (GS-LoRA to GS-LoRA++) | GS-LoRA · CASIA-Face100 4-shot continual 100-20 | H-Mean: 66.02 -> 71.24 | Improved |
| add prototype regularization with full data | GS-LoRA · CASIA-Face100 single-step 100-10 | H-Mean: 71.76 -> 72.04 | No clear effect |
| add group sparse loss (alpha 0 -> 0.01) (ablation) | GS-LoRA++ · CASIA-Face100 4-shot, forget 10 | H-Mean: 66.74 -> 69.47 | Improved |
| raise sparse intensity too far (alpha 0.1 -> 0.5) (ablation) | GS-LoRA++ · CASIA-Face100 4-shot, forget 10 | H-Mean: 71.20 -> 66.98 | Hurt |
| increase LoRA rank from 2 to 8 (ablation) | GS-LoRA++ · Deformable DETR forget 10 classes | H-Mean: 37.98 -> 42.38 | Improved |
| increase LoRA rank from 8 to 16 (ablation) | GS-LoRA++ · Deformable DETR forget 10 classes | H-Mean: 42.38 -> 42.53 | No clear effect |
| increase shots from 1 to 4 (ablation) | GS-LoRA++ · CASIA-Face100 few-shot, forget 10 | H-Mean: 32.14 -> 69.74 | Improved |
| add random augmentation when computing prototypes (1-shot) (ablation) | GS-LoRA++ · CASIA-Face100 1-shot | H-Mean: 32.14 -> 42.64 | Improved |
| add random augmentation when computing prototypes (2-shot) (ablation) | GS-LoRA++ · CASIA-Face100 2-shot | H-Mean: 53.72 -> 57.74 | Improved |
| cut forgetting data ratio from 0.05 to 0.02 (ablation) | GS-LoRA++ · CASIA-Face100 data efficiency | H-Mean: 72.22 -> 67.82 | Hurt |
Closest papers by meaning
- Continual Forgetting for Pre-trained Vision Models (CVPR 2024 · similarity 0.98)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.89)
- Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · similarity 0.88)
- Targeted Forgetting of Image Subgroups in CLIP Models (CVPR 2025 · similarity 0.87)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- Black-Box Forgetting (NeurIPS 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.
