Research map / Architectures and efficiency / RNNs and normalization
GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks
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 |
|---|---|---|---|
| Replace fixed equal task weights with dynamic GradNorm weighting (α=1.5) | VGG SegNet · NYUv2+seg | Depth RMS Err. (m): 0.944 -> 0.925 | Helped |
| Replace fixed equal task weights with dynamic GradNorm weighting (α=1.5) | VGG SegNet · NYUv2+seg | Segmentation Err. (100-IoU): 70.1 -> 67.8 | Helped |
| Replace fixed equal task weights with dynamic GradNorm weighting (α=1.5) | VGG SegNet · NYUv2+seg | Normals Err. (1-|cos|): 0.192 -> 0.174 | Helped |
| Replace equal weighting with GradNorm α=1.5 | Thin ResNet FCN · NYUv2+kpts | Depth RMS Err. (m): 0.697 -> 0.663 | Helped |
| Replace equal weighting with GradNorm α=1.5 | Thin ResNet FCN · NYUv2+kpts | Keypoint Err. (%): 7.80 -> 7.32 | Helped |
| Replace equal weighting with GradNorm α=1.5 | Thin ResNet FCN · NYUv2+kpts | Normals Err. (1-|cos|): 0.172 -> 0.155 | Helped |
| Replace equal weighting with GradNorm α=1.5 | VGG SegNet · NYUv2+kpts | Depth RMS Err. (m): 0.658 -> 0.629 | Helped |
| Replace equal weighting with GradNorm α=1.5 | VGG SegNet · NYUv2+kpts | Keypoint Err. (%): 8.39 -> 7.73 | Helped |
| Replace equal weighting with GradNorm α=1.5 | VGG SegNet · NYUv2+kpts | Normals Err. (1-|cos|): 0.155 -> 0.139 | Helped |
| Tune the asymmetry parameter α toward 1.5 (ablation) | GradNorm · NYUv2+kpts | Average performance gain across tasks | Helped |
| Use larger α to push task weights farther apart (ablation) | GradNorm · NYUv2+kpts | Depth test error: ~5% | Helped |
| Use GradNorm instead of uncertainty weighting | Multitask toy regression · Synthetic 2-task and 10-task systems | Task-normalized test loss | Helped |
Closest papers by meaning
- Just Pick a Sign: Optimizing Deep Multitask Models with Gradient Sign Dropout (NeurIPS 2020 · similarity 0.89)
- Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction (CVPR 2025 · similarity 0.86)
- DiTASK: Multi-Task Fine-Tuning with Diffeomorphic Transformations (CVPR 2025 · similarity 0.86)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.86)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.86)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 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.