Research map / Continual learning / Online and replay-based
Reparameterizing Convolutions for Incremental Multi-Task Learning without Task Interference
Menelaos Kanakis, David Bruggemann, Suman Saha, Stamatios Georgoulis, Anton Obukhov, Luc Van Gool
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 response initialization (RI) (ablation) | RC · PASCAL-Context (5 tasks) | avg per-task performance drop vs single-task (Δm %): 2.13 -> 1.12 | Improved |
| add response initialization (RI) (ablation) | RC+NFF · PASCAL-Context (5 tasks) | avg per-task performance drop vs single-task (Δm %): 2.07 -> 0.99 | Improved |
| add normalized feature fusion (NFF) (ablation) | RC · PASCAL-Context (5 tasks) | avg per-task performance drop vs single-task (Δm %): 2.13 -> 2.07 | Improved |
| add normalized feature fusion (NFF) (ablation) | RC+RI · PASCAL-Context (5 tasks) | avg per-task performance drop vs single-task (Δm %): 1.12 -> 0.99 | Improved |
| replace shared convolutions with reparameterized convolutions (RC) (ablation) | fully shared multi-task model · PASCAL-Context (5 tasks) | avg per-task performance drop vs single-task (Δm %): 3.32 -> 2.13 | Improved |
| make BNs task-specific (ablation) | frozen shared encoder · PASCAL-Context (5 tasks) | avg per-task performance drop vs single-task (Δm %): 14.98 -> 5.76 | Improved |
| make Convs task-specific (ablation) | frozen shared encoder · PASCAL-Context (5 tasks) | avg per-task performance drop vs single-task (Δm %): 14.98 -> 0.62 | Improved |
| use RCM instead of Parallel RA | task-conditional MTL · PASCAL-Context (5 tasks) | avg per-task performance drop vs single-task (Δm %): 2.09 -> 0.99 | Improved |
| use RCM instead of Parallel RA | task-conditional MTL · NYUD (4 tasks) | avg per-task performance drop vs single-task (Δm %): 5.02 -> 1.48 | Improved |
| use RCM instead of ASTMT in incremental setup | incremental learning · PASCAL-Context incremental (low-level tasks first) | avg per-task performance drop vs single-task (Δm %): 11.77 -> 1.26 | Improved |
| use RCM instead of ASTMT in incremental setup (high-level tasks first) | incremental learning · PASCAL-Context incremental (high-level tasks first) | avg per-task performance drop vs single-task (Δm %): 3.76 -> 0.52 | Improved |
| use deeper ResNet-34 backbone instead of ResNet-18 | RCM · PASCAL-Context / NYUD | avg per-task performance drop vs single-task (Δm %): 1.18 and 1.77 | No clear effect |
Closest papers by meaning
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.88)
- Modular Universal Reparameterization: Deep Multi-task Learning Across Diverse Domains (NeurIPS 2019 · similarity 0.88)
- Rectification-Based Knowledge Retention for Task Incremental Learning (TPAMI 2022 · similarity 0.88)
- Task Adaptive Parameter Sharing for Multi-Task Learning (CVPR 2022 · similarity 0.88)
- Rectification-Based Knowledge Retention for Continual Learning (CVPR 2021 · similarity 0.87)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 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.
