Research map / Continual learning / Online and replay-based
CLR: Channel-wise Lightweight Reprogramming for Continual Learning
Yunhao Ge, Yuecheng Li, Shuo Ni, Jiaping Zhao, Ming-Hsuan Yang, Laurent Itti
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 channel-wise lightweight reprogramming with frozen backbone versus best prior dynamic-network method CCLL | CLR · 53-dataset (task-incremental) | average accuracy: 78.3 % -> 86.05 % | Improved |
| use task-specific CLR layers with a shared immutable backbone to isolate parameters | CLR · 53-dataset (task 1 accuracy over time) | forgetting | Improved |
| pretrain backbone on Pascal VOC instead of ImageNet-1k (ablation) | CLR · 53-dataset (task-incremental) | average accuracy: 86.05 % -> 82.49 % | Hurt |
| use self-supervised DINO or SwAV backbone instead of supervised ImageNet-1k (ablation) | CLR · 53-dataset (task-incremental) | average accuracy: 86.05 % -> 85.77 % (DINO), 85.12 % (SwAV) | No clear effect |
| apply CLR layers to all conv kernels including 1x1 (CLR-Full) (ablation) | CLR · 53-dataset (task-incremental) | average accuracy: 86.05 -> 85.85 | No clear effect |
| learn weighted mix of original and reprogrammed feature maps (CLR-mixed) (ablation) | CLR · 53-dataset (task-incremental) | average accuracy: 86.05 -> 86.25 | Improved |
| initialize CLR kernels as identity transformation | CLR · 53-dataset | training efficiency and performance | Improved |
| use CLR instead of Head2Toe for transfer learning | CLR · 53-dataset (transfer learning) | average accuracy: almost 5% average improvement | Improved |
| use replay buffer (ER) instead of plain SGD fine-tuning | SGD baseline · 53-dataset (task-incremental) | average accuracy: 16.71 % -> 53.13 % | Improved |
| apply CLR on incremental CIFAR-100 versus CCLL | CLR · Incremental CIFAR-100 (10 tasks) | average accuracy: 85% -> 94.2% | Improved |
Closest papers by meaning
- CODE-CL: Conceptor-Based Gradient Projection for Deep Continual Learning (ICCV 2025 · similarity 0.90)
- Forget Forgetting: Continual Learning in a World of Abundant Memory (ICLR 2026 · similarity 0.89)
- Convolutional Prompting meets Language Models for Continual Learning (CVPR 2024 · similarity 0.88)
- REP: Resource-Efficient Prompting for Rehearsal-Free Continual Learning (NeurIPS 2025 · similarity 0.88)
- TinySubNets: An Efficient and Low Capacity Continual Learning Strategy (AAAI 2025 · similarity 0.88)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 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.
