Research map / Continual learning / Online and replay-based
Parameter Merging with Gradient-Guided Supermasks in Online Continual Learning
Benliu Qiu, Heqian Qiu, Lanxiao Wang, Taijin Zhao, Yu Dai, Lili Pan, Hongliang Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add gradient-guided supermask parameter merging to ER (ablation) | ER · CIFAR-10 (M=500) | average end accuracy: 56.71 -> 60.98 | Improved |
| add dual-model dual-view distillation on top of supermasks (ablation) | parameter merging · CIFAR-100 (M=1000) | average end accuracy: 28.43 -> 38.35 | Improved |
| add dual-model dual-view distillation on top of supermasks (ablation) | parameter merging · CIFAR-10 (M=500) | learning accuracy: 86.21 -> 85.34 | Hurt |
| vary distillation weight lambda from 4.0 to 7.0 (ablation) | parameter merging · CIFAR-100 (M=1000) | average end accuracy: 38.21 -> 38.36 | No clear effect |
| use 1-order and 2-order supermasks together instead of one alone (ablation) | parameter merging · 10-Split CIFAR-100 | accuracy on current tasks | Improved |
| use Ours instead of MOSE under large memory | parameter merging · CIFAR-100 (M=5000) | average end accuracy: 54.53 -> 52.19 | Hurt |
| use Ours instead of MOSE under small memory | parameter merging · CIFAR-100 (M=1000) | average end accuracy: 35.05 -> 38.35 | Improved |
| use Ours instead of OCM regarding forgetting on CIFAR-10 | parameter merging · CIFAR-10 (M=500) | average forgetting: 13.68 -> 22.33 | Hurt |
Closest papers by meaning
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.89)
- Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of Discretization (CVPR 2025 · similarity 0.88)
- BECAME: Bayesian Continual Learning with Adaptive Model Merging (ICML 2025 · similarity 0.88)
- Rethinking Momentum Knowledge Distillation in Online Continual Learning (ICML 2024 · similarity 0.88)
- Ferret: An Efficient Online Continual Learning Framework under Varying Memory Constraints (CVPR 2025 · similarity 0.88)
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · 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.
