Research map / Reinforcement learning / Policy gradient and actor-critic
Gradient Surgery 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 |
|---|---|---|---|
| Apply PCGrad to MTAN | MTAN · NYUv2 | Within 22.5°: 45.83 -> 46.12 | Helped |
| Apply PCGrad to MTAN | MTAN · NYUv2 | Within 11.25°: 23.17 -> 22.28 | Hurt |
| Add PCGrad to routing networks | Routing, all-fc + WPL · CIFAR-100 | Accuracy: +2.8% absolute test accuracy | Helped |
| Use PCGrad instead of the prior method | CelebA multi-task classifier · CelebA | Average classification error: 8.95 -> 8.69 | Helped |
| Use PCGrad with SAC rather than independent SAC agents | Meta-World MT10 · MT10 | Samples required to solve tasks: about 2 millions fewer samples | Helped |
| Use PCGrad with SAC rather than independent SAC agents | Meta-World MT50 · MT50 | Samples required to solve tasks: about 15 millions fewer samples | Helped |
| Keep the PCGrad-corrected direction but retain the original gradient magnitude (ablation) | PCGrad multi-task RL · Meta-World MT10 and MT50 | Success rate | Hurt |
| Use the PCGrad-computed magnitude but retain the original gradient direction (ablation) | PCGrad multi-task RL · Meta-World MT10 and MT50 | Success rate | Hurt |
| Modify gradient magnitude only, without correcting direction (ablation) | PCGrad multi-task RL · Meta-World MT10 and MT50 | Success rate | Mixed |
| Regularize task-gradient cosine similarities toward 0 instead of projecting only conflicting components (ablation) | Multi-task RL · Meta-World | Average return | Hurt |
| Combine PCGrad with MTAN | NYUv2 multi-task learning · NYUv2 | Validation performance across task metrics: best scores in 8 out of the 9 categories | Helped |
Closest papers by meaning
- Fair Resource Allocation in Multi-Task Learning (ICML 2024 · similarity 0.88)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.88)
- Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction (CVPR 2025 · similarity 0.88)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.88)
- Representation Surgery for Multi-Task Model Merging (ICML 2024 · similarity 0.87)
- Delve into the Applicability of Advanced Optimizers for Multi-Task Learning (NeurIPS 2026 · similarity 0.87)
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.