Research map / Architectures and efficiency / RNNs and normalization
Dynamic Task Prioritization for Multitask Learning
Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, Li Fei-Fei
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 |
|---|---|---|---|
| replace uniform task weights with dynamic task prioritization (hard-first) | hard parameter sharing DenseNet · COCO+MPII multitask (hard sharing, task weighting) | pose AP50: 47.72 -> 56.11 | Improved |
| prioritize hard tasks instead of easy tasks in dynamic priority (ablation) | dynamic task prioritization · COCO+MPII multitask (hard sharing, task weighting) | pose AP50: 32.90 -> 56.11 | Improved |
| use dynamic hard-task priority instead of self-paced weighting | multitask DenseNet · COCO+MPII multitask (hard sharing, task weighting) | pose AP50: 49.34 -> 56.11 | Improved |
| add a learning rate schedule to fixed Prioritize Hard weights (ablation) | Prioritize Hard · COCO+MPII multitask (hard sharing, task weighting) | detection AP50: 65.86 -> 82.64 | Improved |
| add example-level focal loss (gamma=1) to uniform weighting (ablation) | uniform multitask loss · COCO+MPII multitask (hard sharing, task weighting) | segmentation AP50: pose 47.72 -> 54.04; segmentation 78.57 -> 70.77 | Mixed |
| increase task-level focusing parameter from 1 to 2 (ablation) | DTP · COCO+MPII multitask (hard sharing, task weighting) | classification top-1: 91.1 -> 88.8 | Hurt |
| increase example-level focusing parameter gamma0 from 1 to 2 (ablation) | FL · COCO+MPII multitask (hard sharing, task weighting) | detection AP50: 82.0 -> 69.9 | Hurt |
| add example-level focal loss (gamma0=1) on top of DTP (ablation) | DTP · COCO+MPII multitask (hard sharing, task weighting) | detection AP50: detection 79.3 -> 85.9; classification 91.1 -> 88.9 | Mixed |
| replace hard parameter sharing with a task hierarchy ordered by difficulty (ablation) | DenseNet multitask · COCO multitask architecture comparison | pose AP50: 56.11 -> 62.40 | Improved |
| replace UberNet with task hierarchy (ablation) | DenseNet multitask · COCO multitask architecture comparison | segmentation AP50: 77.81 -> 80.26 | Improved |
| place segmentation in the middle of the task hierarchy rather than first (ablation) | task hierarchy · COCO segmentation | relative performance: boosted by 1.25 | Improved |
| train all four tasks jointly instead of single-task detection model | dynamic task prioritization · COCO detection | AP: 49.4 -> 52.5 | Improved |
Closest papers by meaning
- Quantifying Task Priority for Multi-Task Optimization (CVPR 2024 · similarity 0.89)
- Hard Tasks First: Multi-Task Reinforcement Learning Through Task Scheduling (ICML 2024 · similarity 0.88)
- Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.87)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.86)
- Optimal Task Order for Continual Learning of Multiple Tasks (ICML 2025 · similarity 0.86)
- No More Tuning: Prioritized Multi-Task Learning with Lagrangian Differential Multiplier Methods (AAAI 2025 · similarity 0.85)
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.
