Research map / Optimization / Implicit bias and generalization
Beyond Losses Reweighting: Empowering Multi-Task Learning via the Generalization Perspective
Hoang Phan, Lam Tran, Quyen Tran, Ngoc Tran, Tuan Truong, Qi Lei, Nhat Ho, Dinh Phung, Trung Le
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 |
|---|---|---|---|
| add flat-region (sharpness) minimization | CAGrad · MultiFashion | average accuracy: 86.51 -> 87.82 | Improved |
| add flat-region minimization | CAGrad · MultiMNIST | average accuracy: 95.01 -> 95.95 | Improved |
| add flat-region minimization | CAGrad · MultiFashion+MNIST | average accuracy: 92.68 -> 93.54 | Improved |
| add flat-region minimization | MGDA · CelebA | mean error: 9.96 -> 9.22 | Improved |
| add flat-region minimization | CAGrad · CelebA | mean error: 8.52 -> 8.22 | Improved |
| add flat-region minimization | IMTL · NYUv2 | delta m%: -0.76 -> -4.77 | Improved |
| add flat-region minimization | CAGrad · NYUv2 | delta m%: 0.20 -> -3.78 | Improved |
| add flat-region minimization | MGDA · NYUv2 | segmentation mIoU: 30.47 -> 26.42 | Hurt |
| add flat-region minimization | MGDA · CityScapes | delta m%: 44.14 -> 0.67 | Improved |
| add flat-region minimization | IMTL · CityScapes | delta m%: 11.10 -> 1.87 | Improved |
| aggregate loss and flat gradients separately instead of directly aggregating SAM gradients (ablation) | F-MGDA · CityScapes | delta m%: 6.43 -> 0.67 | Improved |
| use single aggregation of loss and flat gradient instead of aggregating each task again (Each-aggre) (ablation) | F-CAGrad · MultiFashion | average accuracy: 87.59 -> 87.82 | No clear effect |
Closest papers by meaning
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.89)
- Delve into the Applicability of Advanced Optimizers for Multi-Task Learning (NeurIPS 2026 · similarity 0.88)
- Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining (NeurIPS 2025 · similarity 0.87)
- Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks (NeurIPS 2025 · similarity 0.87)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.87)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · 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.
