Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Multi-Task Generalization via Regularizing Spurious Correlation
Ziniu Hu, Zhe Zhao, Xinyang Yi, Tiansheng Yao, Lichan Hong, Yizhou Sun, Ed H. Chi
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 MT-CRL with G-IRM Var | MMoE multi-task learning · Multi-MNIST, MovieLens, Taskonomy, CityScape, NYUv2 | average relative performance improvement: +5.5% | Improved |
| Remove G-IRM invariance regularization (ablation) | MT-CRL · Multi-MNIST, MovieLens, Taskonomy, CityScape, NYUv2 | average relative performance improvement: +5.5% -> +3.0% | Hurt |
| Use G-IRM Var instead of G-IRM Norm (ablation) | MT-CRL · Multi-MNIST, MovieLens, Taskonomy, CityScape, NYUv2 | average relative performance improvement: +5.2% -> +5.5% | Improved |
| Add decorrelation regularization (ablation) | MT-CRL · Multi-MNIST | accuracy: 0.882 -> 0.915 | Improved |
| Replace decorrelation regularization with β-VAE (ablation) | MT-CRL · Multi-MNIST | accuracy: 0.915 -> 0.896 | Hurt |
| Add sparsity and balance graph regularization (ablation) | MT-CRL · Multi-MNIST | accuracy: 0.891 -> 0.915 | Improved |
| Vary module count from K=1 to K=8 (ablation) | MT-CRL · Multi-MNIST | accuracy: 0.824 -> 0.915 | Improved |
| Increase module count from K=8 to K=128 (ablation) | MT-CRL · Multi-MNIST | accuracy: 0.915 -> 0.882 | Hurt |
| Train jointly with MTL rather than separately with STL | shared-bottom model · Multi-SEM | validation accuracy: MTL 0.846; STL 0.874 | Hurt |
| Train jointly with MTL rather than separately with STL | shared-bottom model · Multi-MNIST | validation accuracy | Hurt |
Closest papers by meaning
- Identifying and Mitigating Spurious Correlation in Multi-Task Learning (CVPR 2025 · similarity 0.93)
- Generative Causality-Driven Network for Graph Multi-Task Learning (TPAMI 2025 · similarity 0.89)
- CORE-MTL: Rethinking Gradient Balancing via Causal Orthogonal Representations (ICML 2026 · similarity 0.88)
- NTKMTL: Mitigating Task Imbalance in Multi-Task Learning from Neural Tangent Kernel Perspective (NeurIPS 2025 · similarity 0.86)
- MTL-UE: Learning to Learn Nothing for Multi-Task Learning (ICML 2025 · similarity 0.86)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · 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.
