Research map / Optimization / Optimizers
Learning Task-Preferred Inference Routes for Gradient De-Conflict in Multi-Output DNNs
Yi Sun, Xiaochang Hu, Xin Xu, Jian Li, Yifei Shi, Ling‐Li Zeng
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 |
|---|---|---|---|
| train with DR-MGF instead of plain SGD | Vgg-SDN · CIFAR100 | average accuracy: 66.34 -> 69.25 | Improved |
| train with DR-MGF instead of plain SGD | Vgg-SDN · CIFAR10 | average accuracy: 88.15 -> 90.50 | Improved |
| train with DR-MGF instead of plain SGD | Resnet-SDN · CIFAR100 | average accuracy: 59.29 -> 61.84 | Improved |
| train with DR-MGF instead of plain SGD | MSDnet · ImageNet | average accuracy: 66.72 -> 67.35 | Improved |
| add DR-MGF on top of knowledge distillation | MSDnet · CIFAR100 | average accuracy: 72.97 -> 74.13 | Improved |
| use DR-MGF instead of Nash-MTL | SegNet-MTAN · NYUv2 | Delta m%: -4.04 -> -8.39 | Improved |
| apply Nash-MTL instead of plain SGD | Vgg-SDN · CIFAR100 | average accuracy: 66.34 -> 64.29 | Hurt |
| apply Cagrad instead of plain SGD | Vgg-SDN · CIFAR100 exits | accuracy by exit: Exit-1 44.42 -> 53.08; Exit-7 71.76 -> 71.43 | Mixed |
| apply DR-MGF instead of plain SGD at the deepest exit | Resnet-SDN · CIFAR100 Exit-7 | accuracy: 70.02 -> 69.42 | Hurt |
| add network disentanglement to meta-weighted gradient fusion (ablation) | Meta-GF · CIFAR100 | average accuracy: 68.51 -> 69.25 | Improved |
| add meta-weighted gradient fusion to disentanglement with average fusion (ablation) | DR-avgF · CIFAR100 | convergence accuracy | Improved |
| train baselines for 700 epochs to match DR-MGF training time | Cagrad · CIFAR100 Vgg-SDN | average accuracy: 67.60 -> 68.93 | No clear effect |
Closest papers by meaning
- Joint inference and input optimization in equilibrium networks (NeurIPS 2021 · similarity 0.86)
- Quantifying Task Priority for Multi-Task Optimization (CVPR 2024 · similarity 0.85)
- Recon: Reducing Conflicting Gradients From the Root For Multi-Task Learning (ICLR 2023 · similarity 0.85)
- Do Current Multi-Task Optimization Methods in Deep Learning Even Help? (NeurIPS 2022 · similarity 0.84)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.83)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.83)
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.
