Research map / Architectures and efficiency / Neural architecture search
Recon: Reducing Conflicting Gradients From the Root For Multi-Task Learning
Guangyuan SHI, Qimai Li, Wenlong Zhang, Jiaxin Chen, Xiao-Ming Wu
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 |
|---|---|---|---|
| Make layers with the strongest gradient conflicts task-specific | Multi-task learning, including PCGrad · Multi-Fashion+MNIST | T2 accuracy: 88.68 -> 89.77 | Improved |
| Apply Recon to CAGrad | CAGrad · NYUv2 | Δm%: 3.74 -> 5.80 | Improved |
| Apply Recon to Graddrop | Graddrop · CelebA | Average Error: 8.33 -> 8.20 | Improved |
| Apply Recon to BMTAS | BMTAS · PASCAL-Context 4-task | Δm%: -2.89 -> -0.19 | Improved |
| Choose task-specific layers randomly instead of by conflict score (ablation) | Recon · Multi-Fashion+MNIST | severe-conflict reduction: 69.82 -> -52.94 | Hurt |
| Choose a similar number of parameters randomly instead of by conflict score (ablation) | Recon · Multi-Fashion+MNIST | severe-conflict reduction: 69.82 -> 40.13 | Hurt |
| Make the first K layers task-specific rather than conflict-ranked layers (ablation) | Recon with CAGrad · Multi-Fashion+MNIST | T1 accuracy: 98.28 -> 97.63 | Hurt |
| Make the last K layers task-specific rather than conflict-ranked layers (ablation) | Recon with CAGrad · Multi-Fashion+MNIST | T1 accuracy: 98.28 -> 98.21 | No clear effect |
| Select too few layers (ablation) | Recon · Multi-Fashion+MNIST | CAGrad performance | Hurt |
| Select too many layers (ablation) | Recon · Multi-Fashion+MNIST | CAGrad performance | No clear effect |
| Use a high conflict threshold such as S=0.0 (ablation) | Recon · Multi-Fashion+MNIST | CAGrad performance | Hurt |
| Use a very low conflict threshold such as S=-0.15 (ablation) | Recon · Multi-Fashion+MNIST | CAGrad performance | Hurt |
Closest papers by meaning
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.88)
- Quantifying Task Priority for Multi-Task Optimization (CVPR 2024 · similarity 0.87)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.86)
- Fair Resource Allocation in Multi-Task Learning (ICML 2024 · similarity 0.85)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 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.
