Research map / Continual learning / Online and replay-based
Online Curvature-Aware Replay: Leveraging $\mathbf{2^{nd}}$ Order Information for Online Continual Learning
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 OCAR, a second-order replay-based online continual learning optimizer that preconditions gradients with a K-FAC approximation of the Fisher information under an explicit KL stability constraint on buffer data, outperforming prior OCL methods on Split-CIFAR100, Split-TinyImageNet and Online CLEAR.
Problem
Replay-based online continual learning suffers from instability after task shifts (stability gap) and a stability-plasticity tradeoff, without access to task boundaries or identities.
Approach
Formulates replay-based OCL as a second-order joint optimization with a KL-divergence constraint on buffer predictions, and uses the Fisher information (K-FAC approximated, with Tikhonov damping and scheduled lambda and tau) as a gradient preconditioner rather than a quadratic penalty.
Key results
- On Split-CIFAR100 (20 tasks), OCAR reaches 34.9% final accuracy vs 33.3% for LPR, with AAA 48.2 vs 42.5 and WC-Acc 25.0 vs 19.3.
- On Split-TinyImageNet, OCAR gets AAA 38.3 and WC-Acc 17.4, the best among base methods, while ER-ACE has the best final accuracy (23.6).
- On Online CLEAR, OCAR reaches 75.3% accuracy vs 65.2% for LPR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OCAR instead of LPR | Online CLEAR · Online CLEAR (10 Tasks) | AAA: 63.5 \pm 1.0 -> 73.9 \pm 0.5 | Helped |
| Use OCAR instead of LPR | Online CLEAR · Online CLEAR (10 Tasks) | WC-Acc: 62.6 \pm 0.7 -> 70.3 \pm 0.5 | Helped |
| Use OCAR instead of ER-ACE | Split-TinyImageNet · Split-TinyImageNet (20 Tasks) | AAA: 35.0 \pm 1.5 -> 38.3 \pm 1.4 | Helped |
| Use OCAR instead of ER-ACE | Split-TinyImageNet · Split-TinyImageNet (20 Tasks) | Acc: 23.6 \pm 0.7 -> 21.7 \pm 1.0 | Hurt |
| Combine OCAR with ER-ACE (ablation) | OCAR · Split-TinyImageNet (20 Tasks) | AAA: 38.3 \pm 1.4 -> 39.8 \pm 2.0 | Helped |
| Combine OCAR with ER-ACE (ablation) | OCAR · Split-TinyImageNet (20 Tasks) | Probed Acc: 38.3 \pm 0.6 -> 34.7 \pm 0.3 | Hurt |
| Combine OCAR with DER++ (ablation) | OCAR · Split-CIFAR100 (20 Tasks) | WC-Acc: 25.0 \pm 1.1 -> 25.4 \pm 0.8 | Mixed |
| Use OCAR instead of ER | Split MNIST · Split MNIST (5 Tasks) | average accuracy | Helped |
| Increase the ratio of learning rate to Tikhonov regularization | OCAR · Continual learning stream | forgetting | Helped |
Surprising findings
- Combining OCAR with DER++ made optimization much less stable (failing on TinyImageNet), attributed to the entropy regularization breaking the KL/Fisher assumption.
- EWC performed very similarly to plain ER in the online convex setting, while natural gradient was more adaptable but less stable.
Limitations
- Final accuracy (Acc) is noisy and OCAR does not win it on all benchmarks (ER-ACE best on TinyImageNet).
- Hyperparameter selection used only 100 TPE trials and a short four-experience validation stream, which may disadvantage OCAR relative to baselines.
- Justification of the Fisher as preconditioner rests on theory adapted from stationary natural gradient and is not fully established for OCL.
Experimental setup
Datasets: Split-CIFAR100 Split-TinyImageNet Online CLEAR Split MNIST
Benchmarks: Split-CIFAR100 (20 tasks) Split-TinyImageNet (20 tasks) Online CLEAR (10 tasks)
Metrics: Acc AAA WC-Acc Probed Acc
Baselines: ER GDumb AGEM ER+LwF MIR RAR DER++ ER-ACE SCR OnPro OCM LPR
Builds on: experience replay K-FAC natural gradient EWC LPR
Closest papers by meaning
- Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.90)
- Summarizing Stream Data for Memory-Constrained Online Continual Learning (AAAI 2024 · similarity 0.89)
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.89)
- Trust Region Continual Learning as an Implicit Meta-Learner (NeurIPS 2026 · similarity 0.88)
- Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization (CVPR 2025 · similarity 0.88)
- Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling (ICLR 2025 · 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.