Research map / Continual learning / Online and replay-based
Does Continual Learning Equally Forget All Parameters?
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 |
|---|---|---|---|
| finetune only BN (mean&var statistics) after CL, rather than full BN parameters (ablation) | ER on Seq-CIFAR-10 · Seq-CIFAR-10 class-IL buffer 500 | accuracy: 58.90 -> 52.67 | Hurt |
| finetune BN parameters (weights and bias) on buffered data after CL (ablation) | ER on Seq-CIFAR-10 · Seq-CIFAR-10 class-IL buffer 500 | accuracy: 58.90 -> 66.56 (+7.66) | Helped |
| finetune only the bottom conv layer after CL (ablation) | ER on Seq-CIFAR-10 · Seq-CIFAR-10 class-IL buffer 500 | accuracy: 58.90 -> 53.23 | No clear effect |
| finetune last FC layer after CL (ablation) | ER on Seq-CIFAR-10 · Seq-CIFAR-10 class-IL buffer 500 | accuracy: 58.90 -> 66.06 (+7.16) | Helped |
| finetune all parameters on buffer after CL instead of task-specific subset (ablation) | ER on Seq-CIFAR-10 · Seq-CIFAR-10 class-IL buffer 500 | accuracy: 66.34 vs BN 66.56 | No clear effect |
| forgetting prioritized finetuning (FPF) of BN parameters on buffer, small buffer | ER, Seq-CIFAR-10 class-IL buffer 50 vs 5120 · Seq-CIFAR-10 class-IL | accuracy: 33.23 -> 42.42 at buffer 50; 76.61 -> 82.79 at buffer 5120 | Helped |
| finetune all parameters vs task-specific subset as buffer grows (ablation) | ER, Seq-CIFAR-10 class-IL · Seq-CIFAR-10 class-IL buffer 50-5120 | accuracy: subset better at small buffers (50: BN 42.42 vs ALL 40.85); ALL better at 2000+ (78.91 vs BN 78.75; 83.22 vs 82.79) | Mixed |
| apply FPF (BN finetune) to memory-based CL methods | ER, AGEM, FDR, DER, DER++, iCaRL across settings · Seq-MNIST MLP-BN, class-IL buffer 200/500 | accuracy: AGEM 26.40 -> 83.49 (buffer 200) | Helped |
| apply FPF to DER/DER++ on simple settings | DER, DER++ on Seq-MNIST MLP · Seq-MNIST class-IL MLP buffer 500 | accuracy: DER 91.17 -> 91.25; DER++ 91.18 -> 91.22 | No clear effect |
| apply FPF to DER/DER++ on Seq-Tiny-ImageNet with buffer 500 | DER++ on Seq-Tiny-ImageNet ResNet-18 · Seq-Tiny-ImageNet class-IL buffer 500 | accuracy: 8.91 -> 16.02 | Helped |
| finetune all parameters vs BN/FC subset in Tiny-ImageNet at buffer 200 vs 5120 (ablation) | ER, Tiny-ImageNet · Tiny-ImageNet class-IL buffer 200-5120 | accuracy: ALL 11.66 vs BN+FC 12.11 at buffer 200; ALL 36.97 vs BN+FC 36.11 at buffer 5120 | Mixed |
| finetune BN only, applied to CL methods on Seq-PACS domain-IL | ER, DER++ on Seq-PACS · Seq-PACS domain-IL buffer 500 | accuracy: ER 59.07 -> 70.53 (FPF+ER) | Helped |
Closest papers by meaning
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.92)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.91)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.91)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.90)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.90)
- Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective (ICML 2025 · similarity 0.90)
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.