Research map / Continual learning / Online and replay-based
Learning with Preserving for Continual Multitask Learning
Hanchen David Wang, Siwoo Bae, Zirong Chen, Meiyi Ma
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 dynamic weighting to the distance preservation loss (ablation) | LwP · PhysiQ | accuracy (%): 86.0 -> 88.2 | Improved |
| add L_old term (ablation) | LwP · PhysiQ | accuracy (%): 87.1 -> 88.2 | Improved |
| use squared Euclidean distance instead of cosine (ablation) | LwP · PhysiQ | accuracy (%): 85.4 -> 88.2 | Improved |
| use squared Euclidean distance instead of RBF kernel (ablation) | LwP · PhysiQ | accuracy (%): 84.5 -> 88.2 | Improved |
| use DWDP instead of Co2L relational loss (ablation) | LwP · PhysiQ | accuracy (%): 86.4 -> 88.2 | Improved |
| use DWDP instead of RKD (ablation) | LwP · PhysiQ | accuracy (%): 85.1 -> 88.2 | Improved |
| remove dynamic weighting with RKD (ablation) | LwP · PhysiQ | accuracy (%): 85.1 -> 85.9 | Hurt |
| increase input resolution from 32x32 to 224x224 (ResNet50) | LwP · CelebA | accuracy (%): 67.388 -> 85.064 | Improved |
| use LwP instead of strongest CL baseline under combined shift | LwP · BDD100k combined shift | accuracy (%): 76.938 -> 74.004 | Hurt |
| use LwP instead of STL | LwP · FairFace | accuracy (%): 64.435 -> 66.482 | Improved |
| give MTL all labels concurrently instead of CMTL | LwP · PhysiQ | accuracy (%): 88.242 -> 93.536 | Improved |
| add hyperparameter variation (ablation) | LwP · BDD100k weather shift | accuracy (%) | No clear effect |
Closest papers by meaning
- Adversarial Continual Learning (ECCV 2020 · similarity 0.88)
- CLiMB: A Continual Learning Benchmark for Vision-and-Language Tasks (NeurIPS 2022 · similarity 0.88)
- Efficient Continual Learning with Modular Networks and Task-Driven Priors (ICLR 2021 · similarity 0.88)
- HiDe-PET: Continual Learning via Hierarchical Decomposition of Parameter-Efficient Tuning (TPAMI 2025 · similarity 0.88)
- Learning without Isolation: Pathway Protection for Continual Learning (ICML 2025 · similarity 0.87)
- Continual Learning with Global Alignment (NeurIPS 2024 · similarity 0.86)
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.
