Research map / Continual learning / Online and replay-based
Memory-based and Bayesian online learning
Handles blurry or task-free streams with expandable memory, Bayesian updates and meta-learning.
56 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Online and replay-based, keyword cluster “online continual · replay · experience”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 4 of 4 tracked benchmarks.
14 papers in 2023–24 and 12 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 8.6% to 6.1%.
Open questions the data points to
- Split CIFAR10 (FID): best went from 76.58 (up to 2024) to 73.27 in 2025 — Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution (CVPR 2025 · 1 citation)
- Split Fashion (FID): best went from 43.38 (up to 2024) to 37.64 in 2025 — Lifelong Scalable Generative System via Online Maximum Mean Discrepancy (AAAI 2025 · 0 citations)
- Split MNIST (FID): best went from 28.57 (up to 2024) to 18.48 in 2025 — Continual Unsupervised Generative Modelling via Online Optimal Transport (AAAI 2025 · 0 citations)
- Split SVHN (FID): best went from 61.52 (up to 2024) to 51.21 in 2025 — Continual Unsupervised Generative Modelling via Online Optimal Transport (AAAI 2025 · 0 citations)
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Streaming Sparse Gaussian Process Approximations (NeurIPS 2017 · 42 citations)
- 2018 · Variational Continual Learning (ICLR 2018 · 280 citations)
- 2019 · Deep Online Learning Via Meta-Learning: Continual Adaptation for Model-Based RL (ICLR 2019 · 118 citations)
- 2020 · Coresets via Bilevel Optimization for Continual Learning and Streaming (NeurIPS 2020 · 84 citations)
- 2021 · Generalized Variational Continual Learning (ICLR 2021 · 17 citations)
- 2022 · Learning Fast, Learning Slow: A General Continual Learning Method based on Complementary Learning System (ICLR 2022 · 37 citations)
- 2023 · Overcoming Recency Bias of Normalization Statistics in Continual Learning: Balance and Adaptation (NeurIPS 2023 · 3 citations)
- 2024 · Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning (CVPR 2024 · 27 citations)
- 2025 · Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution (CVPR 2025 · 1 citation)
- 2026 · Active Continual Learning with Metaplastic Binary Bayesian Neural Networks (ICML 2026 · 0 citations)
Earliest papers here
- Streaming Sparse Gaussian Process Approximations (NeurIPS 2017 · 42 citations)
- Variational Continual Learning (ICLR 2018 · 280 citations)
- Online Structured Laplace Approximations for Overcoming Catastrophic Forgetting (NeurIPS 2018 · 179 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
Split CIFAR10 · FID (lower is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 76.58 | Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster Memory (CVPR 2024 · 11 citations) |
| 2025 | 73.27 | Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution (CVPR 2025 · 1 citation) |
Split Fashion · FID (lower is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 43.38 | Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster Memory (CVPR 2024 · 11 citations) |
| 2025 | 37.64 | Lifelong Scalable Generative System via Online Maximum Mean Discrepancy (AAAI 2025 · 0 citations) |
Split MNIST · FID (lower is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 28.57 | Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster Memory (CVPR 2024 · 11 citations) |
| 2025 | 18.48 | Continual Unsupervised Generative Modelling via Online Optimal Transport (AAAI 2025 · 0 citations) |
Split SVHN · FID (lower is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 61.52 | Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster Memory (CVPR 2024 · 11 citations) |
| 2025 | 51.21 | Continual Unsupervised Generative Modelling via Online Optimal Transport (AAAI 2025 · 0 citations) |
What has been tried
48 papers here report 543 measured changes (a component added, removed or swapped, compared with a baseline): helped 75%, no clear effect 5%, hurt 17%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| adapt the visual model online while keeping IMU bias as a random walk | EuRoC MH1 · RMSE ATE | 0.093 -> 0.043 | Helped | Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning (CVPR 2024 · 27 citations) |
| Add residual correction | CIFAR-10 Disjoint · AAUC | 77.78±0.52 -> 78.61±0.72 | Helped | Learning Equi-angular Representations for Online Continual Learning (CVPR 2024 · 12 citations) |
| Extend DCM-JS with network expansion (Dynamic-JS) | Split CIFAR10 · accuracy | 49.89± 1.23 -> 55.94 ± 0.58 | Helped | Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster Memory (CVPR 2024 · 11 citations) |
| Add adaptive sampling to reservoir sampling for memory-buffer retention | nuScenes adaptation, M=20 · RA (ADE) | 2.54±0.25 -> 2.45±0.12 | Helped | Continual Learning for Motion Prediction Model via Meta-Representation Learning and Optimal Memory Buffer Retention Strategy (CVPR 2024 · 9 citations) |
| Use fixed class directions e_c instead of estimating spherical means | CIFAR100, M=5k, 10 tasks, retrieval batch 64 · final AA | 25.54±1.48 -> 50.36±0.58 | Helped | Learning Representations on the Unit Sphere: Investigating Angular Gaussian and Von Mises-Fisher Distributions for Online Continual Learning (AAAI 2024 · 6 citations) |
| add class-wise balanced memory (CBM) | CIFAR10 · FAA | 52.60 ±2.43 -> 56.19 ±1.21 | Helped | Dual-Enhanced Coreset Selection with Class-wise Collaboration for Online Blurry Class Incremental Learning (CVPR 2024 · 3 citations) |
| Add multi-level feature fusion using low- and high-level features | Biased datasets average · average accuracy | 43.8 -> 45.5 | Helped | Adaptive Shortcut Debiasing for Online Continual Learning (AAAI 2024 · 2 citations) |
| remove decorrelation | MNIST (10/1) · accuracy | 83.8 (-14.5) | Helped | RanDumb: Random Representations Outperform Online Continually Learned Representations (NeurIPS 2024 · 2 citations) |
| vary the Sinkhorn memory-expansion threshold λ | Split MNIST · FID | No clear effect | Continual Unsupervised Generative Modelling via Online Optimal Transport (AAAI 2025 · 0 citations) | |
| Add memory reduction process | Split MNIST · FID | 33.79 -> 19.23 | Helped | Lifelong Scalable Generative System via Online Maximum Mean Discrepancy (AAAI 2025 · 0 citations) |
| Replace Iter-Fisher with Step-Aware compensation | evaluated settings in Table 4 · online accuracy difference | -56.04±2.78 | Hurt | Ferret: An Efficient Online Continual Learning Framework under Varying Memory Constraints (CVPR 2025 · 0 citations) |
| Combine OCAR with ER-ACE | Split-TinyImageNet (20 Tasks) · AAA | 38.3 ± 1.4 -> 39.8 ± 2.0 | Helped | Online Curvature-Aware Replay: Leveraging $\mathbf{2^{nd}}$ Order Information for Online Continual Learning (ICML 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 10 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR10 · 7 of 10 papers
- MNIST · 5 of 10 papers
- SVHN · 4 of 10 papers
- CIFAR100 · 4 of 10 papers
- Fashion-MNIST · 3 of 10 papers
- CelebA-HQ · 3 of 10 papers
- CACD · 3 of 10 papers
- FFHQ · 3 of 10 papers
Baselines
- CGKD-GAN · 3 of 10 papers
- CGKD-WAE · 3 of 10 papers
- ER · 3 of 10 papers
- R-VAE · 2 of 10 papers
- R-DDPM · 2 of 10 papers
- DER · 2 of 10 papers
- LTS · 2 of 10 papers
- LGM · 2 of 10 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 26 papers; from OpenAlex and the papers' affiliations).
- University of York · 5 papers
- University of Electronic Science and Technology of China · 4 papers
- Korea Advanced Institute of Science and Technology · 3 papers
- Google DeepMind · 2 papers
- Xi’an Jiaotong University · 2 papers
- University of Oxford · 2 papers
Use it
Download BibTeX of 39 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Plug-and-play online training enhancements (39 papers)
- Replay buffer construction and update (33 papers)
Elsewhere in the topic
- Theoretical analysis of forgetting (141 papers)
- Unified and unsupervised continual frameworks (120 papers)
- Gradient projection and plasticity preservation (80 papers)
- Dynamic expansion and isolation (70 papers)
- Plasticity-preserving continual reinforcement learning (43 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).