Research map / Continual learning / Online and replay-based
Efficient Continual Learning with Modular Networks and Task-Driven Priors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use a nearest-neighbor-selected past path as the prior instead of a random past path (ablation) | MNTDP · CTrL direct-transfer and output-shift streams | transfer: -0.2 -> 0; -0.3 -> 0 | Helped |
| Reduce the number of modules by merging adjacent modules (ablation) | MNTDP · CTrL transfer stream and t1^- | transfer: 0 to -0.1; 9% decrease on t1^- | Hurt |
| Increase the number of modules by splitting each module in two (ablation) | MNTDP · CTrL transfer streams | transfer | No clear effect |
| Use MNTDP-D's independent training for candidate paths rather than MNTDP-S's shared-module stochastic training (ablation) | MNTDP · CTrL transfer streams | average accuracy: 0.59 -> 0.64 | Helped |
| Restrict the prior to k=1 rather than considering k=all (ablation) | MNTDP-S · Permuted MNIST and Split CIFAR-100 | average accuracy: 0.97 -> 0.98; 0.75 -> 0.75 | No clear effect |
| Apply the restricted prior when the input distribution changes, despite the search space excluding first-module swaps | MNTDP · CTrL input-shift stream | transfer | Hurt |
| Use modular paths with shared modules across related tasks rather than independent task-specific models | MNTDP-D · CTrL 100-task stream | average accuracy: absolute 18% improvement | Helped |
| Use a modular architecture with an adaptive path prior rather than fixed-capacity methods | MNTDP · CTrL transfer streams | average accuracy: 0.64 vs 0.19 | Helped |
Closest papers by meaning
- Continual Learning via Local Module Composition (NeurIPS 2021 · similarity 0.91)
- Does Continual Learning Equally Forget All Parameters? (ICML 2023 · similarity 0.91)
- A Probabilistic Framework for Modular Continual Learning (ICLR 2024 · similarity 0.91)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.90)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.90)
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · 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.