Research map / Continual learning / Online and replay-based
Compacting, Picking and Growing for Unforgetting Continual Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use CPG rather than PAE, which always picks all old weights (ablation) | continual learning on CIFAR-100 20 tasks · CIFAR-100 20 tasks | average accuracy: 77.1 -> 80.9 | Helped |
| allow critical-weight picking and model expansion rather than forbid expansion as in PackNet (ablation) | continual learning on CIFAR-100 20 tasks · CIFAR-100 20 tasks | average accuracy: 67.5 -> 80.9 | Helped |
| increase the accuracy goal from fine-Avg to fine-Max (ablation) | CPG · CIFAR-100 20 tasks | average accuracy: 80.9 -> 81.2 | Helped |
| increase the accuracy goal to slightly above the fine-tuning maximum (ablation) | CPG · CIFAR-100 20 tasks | average accuracy: 81.2 -> 81.7 | Helped |
| use CPG instead of independent training from scratch | continual learning on CIFAR-100 20 tasks · CIFAR-100 20 tasks (average) | accuracy: 78.8 -> 80.9 | Helped |
| use CPG instead of fine-tuning from a previous task | continual learning on CIFAR-100 20 tasks · CIFAR-100 20 tasks (average) | accuracy: 80.2 -> 80.9 | Helped |
| compress the initial model before subsequent learning | CPG on ImageNet · ImageNet | accuracy: 76.16 -> 75.81 | Hurt |
| reuse a compact accumulated model with CPG rather than fine-tune | fine-grained continual classification · CUBS | accuracy: 82.83 -> 83.59 | Helped |
| reuse a compact accumulated model with CPG rather than fine-tune | fine-grained continual classification · Stanford Cars | accuracy: 91.83 -> 92.80 | Helped |
| reuse a compact accumulated model with CPG rather than fine-tune | fine-grained continual classification · Flowers | accuracy: 96.56 -> 96.62 | Helped |
| reuse a compact accumulated model with CPG rather than fine-tune | fine-grained continual classification · WikiArt | accuracy: 75.60 -> 77.15 | Helped |
| reuse a compact accumulated model with CPG rather than fine-tune | fine-grained continual classification · Sketch | accuracy: 80.78 -> 80.33 | Hurt |
Closest papers by meaning
- Forget Forgetting: Continual Learning in a World of Abundant Memory (ICLR 2026 · similarity 0.89)
- TinySubNets: An Efficient and Low Capacity Continual Learning Strategy (AAAI 2025 · similarity 0.89)
- Task-Free Continual Learning (CVPR 2019 · similarity 0.89)
- Rethinking Continual Learning with Progressive Neural Collapse (ICLR 2026 · similarity 0.89)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.88)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.88)
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.