Research map / Continual learning / Online and replay-based
Overcoming Recency Bias of Normalization Statistics in Continual Learning: Balance and Adaptation
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 |
|---|---|---|---|
| Replace BN with AdaB²N normalization | ER-ACE and DER++ · Online Class-IL on Split CIFAR-10, Split CIFAR-100, and Split Mini-ImageNet | final average accuracy: up to 7.68\%, 6.86\% and 4.26\% | Helped |
| Replace BN with AdaB²N normalization | ER-ACE and DER++ · Online Task-IL and Class-IL benchmarks | final average accuracy | Helped |
| Use AdaB²N instead of BN | ER on ImageNet-sub · ImageNet-sub online Task-IL | accuracy: 72.92 -> 77.24 | Helped |
| Use AdaB²N instead of CN | Online continual learning · DomainNet Domain-IL | accuracy: 10.80 -> 7.81 | Hurt |
| Use AdaB²N instead of BN | Continual learning without replay · Split CIFAR-10 fine-tuning Task-IL | accuracy: 34.97{ \textpm 2.42} -> 36.31{ \textpm 1.72} | Helped |
| Use AdaB²N instead of BN | Joint training · Split CIFAR-10 joint training Task-IL | accuracy: 76.80{ \textpm 1.48} -> 75.79{ \textpm 1.31} | Hurt |
| Vary the training regularization weight λ and testing momentum parameter κ (ablation) | AdaB²N · Online Task-IL and Class-IL | final average accuracy | Mixed |
| Set current-task batch proportion r to 1/10 rather than 1/T (ablation) | AdaB²N · Split CIFAR-10 Task-IL | ACC: 61.61 -> 62.91 | Helped |
| Set current-task batch proportion r to 1/10 rather than 1/T (ablation) | AdaB²N · Split CIFAR-10 Class-IL | ACC: 24.26 -> 25.50 | Helped |
Closest papers by meaning
- Continual Normalization: Rethinking Batch Normalization for Online Continual Learning (ICLR 2022 · similarity 0.92)
- Rebalancing Batch Normalization for Exemplar-Based Class-Incremental Learning (CVPR 2023 · similarity 0.90)
- NoTA: Normalized Tensor Adaptation for Parameter-Efficient Continual Learning (NeurIPS 2026 · similarity 0.89)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.89)
- A Statistical Theory of Regularization-Based Continual Learning (ICML 2024 · similarity 0.88)
- Confounder-Free Continual Learning via Recursive Feature Normalization (ICML 2025 · 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.