Research map / Continual learning / Online and replay-based
Perturbing to Preserve: Defending Fragile Knowledge in Online Continual Learning
Dulan Zhou, Zijian Gao, Kele Xu
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 EMA teacher (temporal defense) to ER (ablation) | ER · CIFAR-100 (M=1k) | accuracy: 25.12 -> 37.50 | Improved |
| add spatial perturbation with consistency to ER (ablation) | ER · CIFAR-100 (M=1k) | accuracy: 25.12 -> 35.45 | Improved |
| add spatial perturbation to ER + EMA (full method) (ablation) | ER + EMA · CIFAR-100 (M=1k) | accuracy: 37.50 -> 39.87 | Improved |
| add EMA to ER (ablation) | ER · CIFAR-100 (M=1k) | average forgetting: 12.42 -> 9.65 | Improved |
| add spatial perturbation to ER + EMA (ablation) | ER + EMA · CIFAR-100 (M=1k) | average forgetting: 9.65 -> 8.91 | Improved |
| add Gaussian-noise perturbation on all buffer samples (ablation) | PDFK · CIFAR-100 (M=1k) | accuracy: 36.61 -> 37.42 | Improved |
| use maximal perturbation instead of Gaussian noise (all samples) (ablation) | PDFK · CIFAR-100 (M=1k) | accuracy: 37.42 -> 38.65 | Improved |
| restrict perturbation to correctly classified samples x* (ablation) | PDFK · CIFAR-100 (M=1k) | accuracy: 38.65 -> 39.87 | Improved |
| restrict perturbation to correctly classified samples x* (ablation) | PDFK · CIFAR-100 (M=1k) | average forgetting: 9.14 -> 8.91 | Improved |
| PDFK on top of MKD: parameter-perturbation sensitivity | MKD · parameter perturbation | Delta Acc AUC: 0.05101 -> 0.03500 | Improved |
| PDFK on top of MKD: initial degradation slope | MKD · parameter perturbation | Delta Acc Slope0: 6.89 -> 2.97 | Improved |
| PDFK over ER + MKD, blurry Tiny-ImageNet M=5000 | ER + MKD · Tiny-ImageNet blurry | accuracy: 31.80 -> 34.51 | Improved |
Closest papers by meaning
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.88)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.88)
- Spurious Forgetting in Continual Learning of Language Models (ICLR 2025 · similarity 0.88)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.88)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.87)
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 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.
