Research map / Continual learning / Online and replay-based
Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization
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.
Reformulates replay-based continual learning as hierarchical gradient aggregation and proposes POCL, which weights previous-task gradients via Pareto (min-norm Frank-Wolfe) optimization with hyper-gradients, improving online class-incremental accuracy on Split CIFAR-10/100 and TinyImageNet.
Problem
Catastrophic forgetting in online class-incremental continual learning, where replay methods weight previous-task gradients heuristically and ignore interrelations among previous tasks, yielding sub-optimal past-task performance.
Approach
Unifies experience replay and gradient alignment as a hierarchical gradient aggregation framework, then chooses the aggregation weights over previous-task gradients by solving a min-norm problem (Pareto optimality over tasks) and implements gradients as hyper-gradients to avoid conflict with the current task gradient.
Key results
- On Split CIFAR-10 with buffer 1k, POCL reaches 58.50% Acc and 66.23% AAA, versus 52.08% Acc for the next-best CBA.
- On Split CIFAR-100 with buffer 5k, POCL reaches 33.36% Acc versus 20.58% for ER-ACE.
- Ablation on Split CIFAR-10 shows the hyper-gradient implementation contributes most of the gain (Acc 19.45 to 54.57), with Pareto weighting adding a further gain (54.57 to 58.50).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use POCL instead of CBA | online class-incremental learning · Split CIFAR-10 (|M|=1k) | AAA: 65.47 -> 66.23 | Helped |
| Use POCL instead of CLSER | online class-incremental learning · Split CIFAR-10 (|M|=1k) | Acc: 53.06 -> 58.50 | Helped |
| Use POCL instead of ER-ACE | online class-incremental learning · Split CIFAR-100 (|M|=5k) | Acc: 20.58 -> 33.36 | Helped |
| Use POCL instead of ER-ACE | online class-incremental learning · Split TinyImageNet (|M|=2k) | Acc: 13.23 -> 12.69 | Hurt |
| Add the hyper-gradient implementation (ablation) | POCL with Pareto optimization · Split CIFAR-10 (|M|=1k) | AAA: 39.03 -> 66.23 | Helped |
| Add the hyper-gradient implementation (ablation) | POCL with Pareto optimization · Split CIFAR-100 (|M|=5k) | Acc: 11.34 -> 33.36 | Helped |
| Add Pareto optimization (ablation) | POCL with hyper-gradient · Split CIFAR-10 (|M|=1k) | Acc: 54.57 -> 58.50 | Helped |
| Add Pareto optimization (ablation) | POCL with hyper-gradient · Split CIFAR-100 (|M|=5k) | AAA: 35.89 -> 36.34 | Helped |
Limitations
- Some baselines (MER, ER-ACE on certain settings) are implemented as time-consuming or unstable and reported as '-' in tables.
- FM comparisons with ER-ACE and La-MAML are deemed unfair due to differing design focus, so those are deferred to the appendix.
Experimental setup
Datasets: Split CIFAR-10 Split CIFAR-100 Split TinyImageNet
Benchmarks: Split CIFAR-10 Split CIFAR-100 Split TinyImageNet
Metrics: Average Accuracy Forgetting Measure Anytime Average Accuracy
Baselines: SGD EWC GEM AGEM ER MER La-MAML DER DER++ ER-ACE CLSER CBA
Builds on: GEM MER La-MAML ER Frank-Wolfe
Closest papers by meaning
- Pareto Continual Learning: Preference-Conditioned Learning and Adaption for Dynamic Stability-Plasticity Trade-off (AAAI 2025 · similarity 0.93)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.91)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.90)
- Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.90)
- Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective (ICML 2025 · similarity 0.89)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.89)
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.