Research map / Architectures and efficiency / RNNs and normalization
Efficiently Identifying Task Groupings for 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 groups by maximizing measured inter-task gradient affinity | multi-task learning on Taskonomy · Taskonomy, 2-split grouping | total test loss: 0.5940 -> 0.5288 | Helped |
| select task groups by maximizing measured inter-task gradient affinity | multi-task learning on CelebA · CelebA, 9-task subset, 2-split grouping | total error: 50.00 -> 49.66 | Helped |
| measure inter-task affinity every 5 rather than every training step (ablation) | TAG · CelebA, 9-task subset | relative performance: 5.13 -> 5.13 | No clear effect |
| measure inter-task affinity every 10 rather than every training step (ablation) | TAG · CelebA, 9-task subset | relative performance: 5.13 -> 5.13; 3.19x | No clear effect |
| measure inter-task affinity every 25 rather than every training step (ablation) | TAG · CelebA, 9-task subset | relative performance: 5.13 -> 4.84 | Hurt |
| measure inter-task affinity every 50 rather than every training step (ablation) | TAG · CelebA, 9-task subset | relative performance: 5.13 -> 3.73 | Hurt |
| measure inter-task affinity every 100 rather than every training step (ablation) | TAG · CelebA, 9-task subset | relative performance: 5.13 -> 2.06 | Hurt |
| use only the first 25% of training to compute affinities (ablation) | TAG · CelebA, 9-task subset | relative performance: 5.13 -> 3.67 | Hurt |
| use only the middle 25% of training to compute affinities (ablation) | TAG · CelebA, 9-task subset | relative performance: 5.13 -> 4.31 | Hurt |
| use only the final 25% of training to compute affinities (ablation) | TAG · CelebA, 9-task subset | relative performance: 5.13 -> 4.02 | Hurt |
| compute affinities from validation rather than training loss (ablation) | TAG · CelebA, 9-task subset | average total error: 49.576 vs 49.574 | No clear effect |
| add PCGrad to TAG-selected groups (ablation) | TAG · CelebA, 2-, 3-, and 4-split groupings | performance change: 0.85%; 0.18%; -0.08% | Mixed |
Closest papers by meaning
- Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.93)
- DMTG: One-Shot Differentiable Multi-Task Grouping (ICML 2024 · similarity 0.89)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.87)
- Task Groupings Regularization: Data-Free Meta-Learning with Heterogeneous Pre-trained Models (ICML 2024 · similarity 0.86)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.86)
- Joint-Task Regularization for Partially Labeled Multi-Task Learning (CVPR 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.