Research map / Label-efficient and robust learning / Domain adaptation and generalization
Scalarization for Multi-Task and Multi-Domain Learning at Scale
Amelie Royer, Tijmen Blankevoort, Babak Ehteshami Bejnordi
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 |
|---|---|---|---|
| Tune scalarization weights with PBT instead of using uniform weights | DomainNet joint training with ResNet26 at 0.25 width · DomainNet, 6 domains | average accuracy: 46.78 ± 0.10 -> 48.01 ± 0.08 | Improved |
| Tune scalarization weights with PBT instead of using uniform weights | DomainNet joint training with ResNet26 at original width · DomainNet, 6 domains | average accuracy: 48.83 ± 0.08 -> 49.27 ± 0.12 | Improved |
| Tune scalarization weights with PBT instead of using uniform weights | DomainNet joint training with ResNet50 at 0.25 width · DomainNet, 6 domains | average accuracy: 49.69 ± 0.05 -> 50.69 ± 0.10 | Improved |
| Tune scalarization weights with PBT instead of using uniform weights | DomainNet joint training with ResNet50 at original width · DomainNet, 6 domains | average accuracy: 51.53 ± 0.06 -> 51.83 ± 0.06 | Improved |
| Use PB2-tuned scalarization weights rather than uniform weights | CelebA 8-task MTL with ViT-S/4, 6 layers, full width · CelebA | average accuracy: 91.23 -> 91.28 | No clear effect |
| Use PB2-tuned scalarization weights rather than uniform weights | CelebA 8-task MTL with ViT-S/4, 9 layers, full width · CelebA | average accuracy: 91.17 -> 91.25 | No clear effect |
| Increase model capacity for joint multi-task or multi-domain training | Scalarized MTL/MDL · CIFAR/STL, DomainNet, CelebA, and Taskonomy comparisons with single-dataset models | test performance relative to single-dataset training | Improved |
| Tune task/domain scalarization weights rather than defaulting to uniform weights | Scalarized MTL/MDL · High-imbalance task/domain combinations | test performance | Improved |
| Use a smaller model in the same architecture family to search scalarization weights before scaling up | Scalarized MTL/MDL · Taskonomy and DomainNet | relative performance ranking across weights | No clear effect |
| Correct conflicting gradients during every training iteration | Uniform scalarization MTL/MDL · CelebA and DomainNet gradient-conflict analysis | final model performance | No clear effect |
| Use uniform weights instead of PBT-tuned weights | DomainNet joint training with ResNet26 at 0.25 width · DomainNet, infograph domain | accuracy: 19.45 ± 0.08 -> 20.46 ± 0.15 | Improved |
| Use PBT-tuned weights instead of uniform weights | DomainNet joint training with ResNet26 at 0.25 width · DomainNet, real domain | accuracy: 55.64 ± 0.01 -> 60.43 ± 0.05 | Improved |
Closest papers by meaning
- PiKE: Adaptive Data Mixing for Large-Scale Multi-Task Learning Under Low Gradient Conflicts (NeurIPS 2025 · similarity 0.87)
- Scaling Laws for Optimal Data Mixtures (NeurIPS 2025 · similarity 0.86)
- GRAPE: Optimize Data Mixture for Group Robust Multi-target Adaptive Pretraining (NeurIPS 2025 · similarity 0.85)
- Inductive biases of multi-task learning and finetuning: multiple regimes of feature reuse (NeurIPS 2024 · similarity 0.85)
- Task Groupings Regularization: Data-Free Meta-Learning with Heterogeneous Pre-trained Models (ICML 2024 · similarity 0.85)
- Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation (ICML 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.
