Research map / Architectures and efficiency / RNNs and normalization
Which Tasks Should Be Learned Together in Multi-task Learning?
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 |
|---|---|---|---|
| Select task groupings computationally rather than using all-in-one or independent networks | Task grouping framework · Task Set 1, Setting 1 | overall test loss | Helped |
| Use task groupings selected by the Higher Order Approximation instead of fully trained optimal groupings (ablation) | Task grouping framework · All four settings | overall test loss | No clear effect |
| Use early-stopped candidate networks to select groups, then train chosen networks to convergence (ablation) | Task grouping framework · All four settings | overall test loss | Mixed |
| Predict higher-order network performance from pairwise networks (ablation) | Task grouping framework · Task Set 1, Setting 1 | training time: 45% | Helped |
| Use early stopping for candidate-network evaluation (ablation) | Task grouping framework · Task Set 1, Setting 1 | training time: 20x | Helped |
| Train Surface Normal Prediction jointly with other tasks | Task Set 1 multi-task models · Task Set 1, Settings 1–3 | relative performance of co-trained task: 15 out of 16 models | Helped |
| Train task pairs jointly instead of separately | Task Set 1, Setting 1 (Xception17) · Task Set 1, Setting 1 | relative performance (%): -1.15 | Hurt |
| Train task pairs jointly instead of separately | Task Set 1, Setting 2 (Xception) · Task Set 1, Setting 2 | relative performance (%): 13.54 | Helped |
| Train task pairs jointly instead of separately | Task Set 1, Setting 3 (200,000 training instances) · Task Set 1, Setting 3 | relative performance (%): 2.6 | Mixed |
| Train Auto Encoder with other Task Set 2 tasks | Task Set 2 multi-task models · Task Set 2, Setting 4 | relative performance of co-trained tasks | Hurt |
| Use higher-capacity networks rather than Xception17 | Pairwise multi-task training · Task Set 1, Setting 2 versus Setting 1 | pairwise task affinity | Helped |
| Use fewer training examples rather than the full training set | Pairwise multi-task training · Task Set 1, Setting 3 versus Setting 2 | pairwise task affinity | Hurt |
Closest papers by meaning
- One Network Fits All? Modular versus Monolithic Task Formulations in Neural Networks (ICLR 2021 · similarity 0.90)
- Efficiently Identifying Task Groupings for Multi-Task Learning (NeurIPS 2021 · similarity 0.89)
- Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.87)
- Quantifying Task Priority for Multi-Task Optimization (CVPR 2024 · similarity 0.87)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.85)
- Provable Multi-Task Representation Learning by Two-Layer ReLU Neural Networks (ICML 2024 · 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.