Research map / Continual learning / Online and replay-based
Pareto Continual Learning: Preference-Conditioned Learning and Adaption for Dynamic Stability-Plasticity Trade-off
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 experience-replay continual learning's stability-plasticity trade-off as multi-objective optimization and trains a hypernetwork-conditioned classifier over sampled preference vectors, selecting the lowest-entropy preference output at inference.
Problem
Experience replay methods in class-incremental continual learning use a fixed stability-plasticity trade-off, which ignores that the optimal trade-off changes over training and inference.
Approach
Treat the replay loss (stability) and new-data loss (plasticity) as two MOO objectives and learn a single preference-conditioned model via a hypernetwork that generates the final linear layer from a sampled preference vector; at inference, per-sample selection of the preference giving the lowest-entropy output provides dynamic trade-off adaptation.
Key results
- ParetoCL reaches 59.95% final accuracy on Seq-CIFAR10 (online), versus 50.33% for DER++ and 58.43% for VR-MCL.
- On Seq-CIFAR100 (online), ParetoCL achieves 24.45% Acc and 33.04 AAA, outperforming VR-MCL's 22.31% Acc and 30.46 AAA.
- Training time on Seq-CIFAR10 is 224.62 s for ParetoCL versus 1073.61 s for VR-MCL, roughly one-fifth, which the paper describes as less than one-fourth.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use preference-conditioned learning and dynamic preference adaptation instead of VR-MCL | Continual learning with ResNet-18 · Seq-CIFAR10 online | Acc: 58.43 -> 59.95 | Helped |
| Use preference-conditioned learning and dynamic preference adaptation instead of VR-MCL | Continual learning with ResNet-18 · Seq-CIFAR100 online | Acc: 22.31 -> 24.45 | Hurt |
| Use preference-conditioned learning and dynamic preference adaptation instead of VR-MCL | Continual learning with ResNet-18 · Seq-TinyImageNet online | Acc: 20.55 -> 23.09 | Helped |
| Use preference-conditioned learning and dynamic preference adaptation instead of VR-MCL | Continual learning with ResNet-18 · Seq-CIFAR10 offline | Acc: 68.14 -> 69.55 | Helped |
| Use preference-conditioned learning and dynamic preference adaptation instead of VR-MCL | Continual learning with ResNet-18 · Seq-TinyImageNet offline | Acc: 28.02 -> 28.28 | Helped |
| Use dynamic preference adaptation rather than fixed inference preference (0.5, 0.5) (ablation) | ParetoCL · Seq-CIFAR100 online | average incremental accuracy | Helped |
| Increase buffer size from M=600 to M=1000 (ablation) | ParetoCL · Seq-CIFAR100 online | Acc: 20.49 ±0.29 -> 24.45 ±0.39 | Helped |
| Increase buffer size from M=1000 to M=1400 (ablation) | ParetoCL · Seq-CIFAR100 online | Acc: 24.45 ±0.39 -> 26.97 ±0.42 | Helped |
Surprising findings
- A static-preference variant (ParetoCL--) still performs competitively with other baselines, suggesting that preference-conditioned training itself improves the feature extractor.
- Most learned Pareto solutions cluster in a balanced corner region rather than at the extremes.
Limitations
- Relies on a hypernetwork to generate classifier parameters; more sophisticated conditioning may help.
- Inference cost scales with the number of sampled preference vectors (K=20), and the method is evaluated only on class-incremental image benchmarks with ResNet-18.
Experimental setup
Datasets: Seq-CIFAR10 Seq-CIFAR100 Seq-TinyImageNet
Benchmarks: Seq-CIFAR10 Seq-CIFAR100 Seq-TinyImageNet
Metrics: Acc AAA (averaged anytime accuracy)
Baselines: SGD OnEWC LwF SI La-MAML VR-MCL ER A-GEM GEM DER DER++ CLSER
Builds on: experience replay DER++ Pareto multi-task learning MGDA Tchebycheff scalarization
Closest papers by meaning
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.93)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.89)
- BECAME: Bayesian Continual Learning with Adaptive Model Merging (ICML 2025 · similarity 0.89)
- Temporal-Difference Variational Continual Learning (NeurIPS 2025 · similarity 0.89)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 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.