Research map / Architectures and efficiency / RNNs and normalization
In Defense of the Unitary Scalarization for Deep Multi-Task Learning
Vitaly Kurin, Alessandro De Palma, Ilya Kostrikov, Shimon Whiteson, M. Pawan Kumar
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 specialized multi-task optimizers (MGDA, IMTL, PCGrad, GradDrop, RLW) with unitary scalarization plus regularization | Multi-MNIST encoder-decoder · Multi-MNIST | avg task test accuracy | No clear effect |
| unitary scalarization with l2 and dropout instead of specialized optimizers | CelebA 40-task · CelebA | avg task test accuracy | No clear effect |
| unitary scalarization instead of specialized optimizers | Cityscapes · Cityscapes | depth error, segmentation mIOU | No clear effect |
| stabilize unitary scalarization with larger replay buffer, actor l2, reward normalization at sampling | SAC on Meta-World MT10 · Meta-World MT10 | best average success rate | No clear effect |
| use PCGrad instead of unitary scalarization | SAC on Meta-World MT50 · Meta-World MT50 | best average success rate | No clear effect |
| add explicit regularization (dropout, l2) to unitary scalarization (ablation) | unitary scalarization · CelebA | peak avg validation accuracy | Improved |
| use specialized optimizers (without explicit regularization) instead of unregularized unitary scalarization (ablation) | CelebA · CelebA validation | overfitting onset | Improved |
| use PCGrad instead of unitary scalarization | CelebA · CelebA | training time per epoch: roughly 35 times slower | Hurt |
Closest papers by meaning
- Delve into the Applicability of Advanced Optimizers for Multi-Task Learning (NeurIPS 2026 · similarity 0.88)
- Fair Resource Allocation in Multi-Task Learning (ICML 2024 · similarity 0.86)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.86)
- Representation Learning Enables Scalable Multitask Deep Reinforcement Learning (NeurIPS 2026 · similarity 0.86)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.85)
- Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction (CVPR 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.
