Research map / Continual learning / Continual fine-tuning of LLMs
Learning with Recoverable Forgetting
Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, Xinchao Wang
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 pruning-based initialisation of the deposit module (ablation) | LIRF · CIFAR-10 | Avg Acc (recover net): 95.15 -> 95.49 | Improved |
| add attention loss Lat to the knowledge-preservation term (ablation) | LIRF · CIFAR-10 | Avg Acc (recover net): 95.09 -> 95.15 | No clear effect |
| add attention loss Lat to Lkr (ablation) | LIRF target net · CIFAR-10, ResNet18 student, AT-based distillation | Dep Acc of distilled student: 94.61 -> 88.76 | Improved |
| add attention loss Lat to Lkr (ablation) | LIRF target net · CIFAR-10, CNN student, AT-based distillation | Dep Acc of distilled student: 85.72 -> 81.83 | Improved |
| add attention loss Lat to Lkr (ablation) | LIRF target net · CIFAR-10, ResNet18 student, logit-based distillation | Dep Acc of distilled student: 92.70 -> 91.49 | Improved |
| add attention loss Lat to Lkr (ablation) | LIRF target net · CIFAR-10, CNN student, logit-based distillation | Dep Acc of distilled student: 84.26 -> 82.75 | Improved |
| add Lpt and Lre to Lkr+Lkp (w/o Lat) (ablation) | LIRF · CIFAR-10 | Avg Acc (recover net): - -> 95.09 | Improved |
| replace incremental-learning-style training with LIRF deposit/withdraw | recover net · CIFAR-10 | Avg Acc: 93.12 -> 95.49 | Improved |
| use LIRF recover net instead of original network | ResNet-18 · CIFAR-10 | Avg Acc: 94.06 -> 95.57 | Improved |
| use LIRF recover net instead of original network | ResNet-18 · CUB200-2011 | Avg Acc: 49.81 -> 54.01 | Improved |
| increase division block number n (ablation) | LIRF · CIFAR-10 | Dep Acc and Avg Acc | Hurt |
| increase pruning rate towards 50% (ablation) | LIRF deposit module · CIFAR-10 | Avg Acc (recover net) | Mixed |
Closest papers by meaning
- A Comprehensive Survey of Forgetting in Deep Learning Beyond Continual Learning (TPAMI 2024 · similarity 0.89)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.87)
- Fast Machine Unlearning without Retraining through Selective Synaptic Dampening (AAAI 2024 · similarity 0.87)
- Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · similarity 0.87)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.87)
- Learning without Isolation: Pathway Protection for Continual Learning (ICML 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.
