Research map / Continual learning / Online and replay-based
Federated Continual Learning Goes Online: Uncertainty-Aware Memory Management for Vision Tasks and Beyond
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes O-FCL, an online federated continual learning method that keeps a per-client memory buffer chosen by a Bregman-Information (test-time augmentation based) epistemic uncertainty score, evaluated on vision, biomedical and text streams.
Problem
Federated continual learning assumes offline, multi-epoch training with generative replay, which fails when clients receive data as single-pass mini-batch streams and is mostly limited to images.
Approach
Formalises online FCL and uses an uncertainty-aware memory: samples are stored per class by a Bregman Information variance term (from bias-variance decomposition of cross-entropy, estimated via TTA in logit space), with random replay and burn-in plus periodic averaging on the server.
Key results
- On CIFAR10 (M=1000), BI with bottom-k selection reaches 35.83% last accuracy and 19.07% forgetting, versus 33.64% (ER) and FedAvg at 16.90%.
- On CRC-Tissue (M=120), BI achieves 62.33% accuracy and 7.99% forgetting, the best among compared methods in that table.
- On text datasets, BI attains the best accuracy/forgetting on 20NewsGroups (46.72%, 26.97%), DBPedia (78.68%, 21.79%) and Yahoo Answers (50.13%, 32.65%) at the reported memory sizes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BI uncertainty-based memory management instead of standard FL or generative replay | O-FCL · CIFAR10 | average last accuracy: 16.90 -> 35.83 | Helped |
| Use BI uncertainty-based memory management instead of standard FL or generative replay | O-FCL · CIFAR100 | average last accuracy: 4.60 -> 14.31 | Helped |
| Use BI uncertainty-based memory management instead of standard FL or generative replay | O-FCL · CRC-Tissue | average last accuracy: 27.62 -> 62.33 | Helped |
| Use BI uncertainty-based memory management instead of standard FL or generative replay | O-FCL · KC-Cell | average last accuracy: 16.32 -> 21.61 | Helped |
| Select bottom-k least-uncertain, class-representative samples rather than top-k high-uncertainty samples (ablation) | BI memory management · CIFAR10 | average last accuracy: 21.31 ± 1.46 -> 24.89 ± 0.83; 26.57 ± 0.83 -> 27.84 ± 2.31; 35.12 ± 2.51 -> 35.83 ± 2.60 | Mixed |
| Increase burn-in period from 1 to 10 (ablation) | O-FCL · CIFAR10 | accuracy: 23.77 -> 25.64 | Helped |
| Increase burn-in period from 1 to 30 (ablation) | O-FCL · CIFAR10 | accuracy: 23.77 -> 24.89 | Helped |
| Change the parameter averaging strategy from FedAvg to Weighted FedAvg (ablation) | O-FCL · CIFAR10 | forgetting: 44.38 -> 37.77 | Helped |
| Set jump parameter q to 3 instead of q=5 (ablation) | O-FCL · CIFAR10 | accuracy: 24.89 -> 37.83 | Helped |
| Reduce augmentation set from 12 to 10 (ablation) | BI uncertainty estimation · CIFAR10 | accuracy: 24.89 -> 22.18 | Hurt |
| Increase task count from 5 to 25 (ablation) | BI (ours) · CIFAR100 | average last forgetting: 8.96 -> 33.65 | Hurt |
Surprising findings
- Generative FCL baselines (MFCL, FedCIL) underperform simple memory replay in the online setting, with MFCL's synthetic images showing no informative patterns.
- On KC-Cell the BI method has very high forgetting (59.05%) relative to its accuracy of 21.61%, and ER attains the higher accuracy at M=160 (22.29%).
Limitations
- BI estimation requires test-time augmentation ensembling, reducing efficiency relative to ER.
- The BI estimator is only asymptotically unbiased and underestimates the theoretical quantity.
- Results on several datasets show high variance and mixed rankings, with the best setting varying across memory sizes and selection strategies.
Experimental setup
Datasets: CIFAR10 CIFAR100 TinyImageNet CRC-Tissue KC-Cell 20NewsGroups DBPedia Yahoo Answers
Benchmarks: CIFAR10 CIFAR100 TinyImageNet CRC-Tissue KC-Cell 20NewsGroups DBPedia Yahoo Answers
Metrics: average last accuracy average last forgetting
Baselines: FedAvg FedProx Weighted FedAvg ER CBR MFCL FedCIL least confidence margin sampling ratio of confidence entropy
Builds on: Experience Replay FedAvg FedProx bias-variance decomposition of cross-entropy class-balanced reservoir (CBR)
Closest papers by meaning
- Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.92)
- Federated Orthogonal Training: Mitigating Global Catastrophic Forgetting in Continual Federated Learning (ICLR 2024 · similarity 0.89)
- Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution (CVPR 2025 · similarity 0.88)
- Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online Continual Learning (NeurIPS 2024 · similarity 0.88)
- CAN: Leveraging Clients As Navigators for Generative Replay in Federated Continual Learning (ICML 2025 · similarity 0.88)
- C$^2$Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning (NeurIPS 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.