Research map / Learning theory and algorithms / Prediction and decision losses
Learning Multiple Pixelwise Tasks Based on Loss Scale Balancing
Jae-Han Lee, Chul Lee, Chang-Su Kim
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 |
|---|---|---|---|
| use LSB loss scale balancing instead of equal weighting | multi-column MTL, MobileNet.v2 · NYUv2 | average rank: 3.92 -> 1.00 | Improved |
| use LSB instead of equal weighting | full-sharing MTL, MobileNet.v2 · NYUv2 | average rank: 4.00 -> 1.50 | Improved |
| use LSB instead of equal weighting | multi-decoder MTL, MobileNet.v2 · NYUv2 | average rank: 2.92 -> 2.25 | Improved |
| use LSB instead of equal weighting | full-sharing MTL, PNASNet · NYUv2 | average rank: 3.17 -> 1.33 | Improved |
| use LSB instead of equal weighting | multi-column MTL, PNASNet · NYUv2 | average rank: 4.00 -> 1.33 | Improved |
| use LSB instead of equal weighting | multi-column MTL on Taskonomy 8-task · Taskonomy | average rank: 3.56 -> 1.88 | Improved |
| use LSB instead of GLS for 2D edge loss | multi-column MTL on Taskonomy 8-task · Taskonomy | 2D edge loss: 0.119 -> 0.114 | Improved |
| use LSB instead of equal weighting for landmark localization | multi-column MTL · CelebA | eye pixel distance: 0.859 -> 0.641 | Improved |
| use LSB instead of equal weighting for attribute classification | multi-column MTL · CelebA | accuracy: 90.6% -> 90.8% | No clear effect |
| increase beta gradually from 0 (curriculum on task difficulty) instead of ignoring difficulty (beta = 0) (ablation) | LSB · NYUv2 segmentation | mIoU: 27.5% -> 32.6% | Improved |
| use fixed beta = 1 (focus on difficult tasks) instead of beta = 0 (ablation) | LSB · NYUv2 segmentation | mIoU: 27.5% -> 28.1% | Improved |
| use negative beta = -1 (focus on easy tasks) instead of beta = 0 (ablation) | LSB · NYUv2 segmentation | mIoU: 27.5% -> 25.5% | Hurt |
Closest papers by meaning
- Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.85)
- Transforming Vision Transformer: Towards Efficient Multi-Task Asynchronous Learner (NeurIPS 2024 · similarity 0.85)
- TADFormer: Task-Adaptive Dynamic TransFormer for Efficient Multi-Task Learning (CVPR 2025 · similarity 0.85)
- Robust Multi-Task Learning with Excess Risks (ICML 2024 · similarity 0.83)
- DiTASK: Multi-Task Fine-Tuning with Diffeomorphic Transformations (CVPR 2025 · similarity 0.83)
- Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining (NeurIPS 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.
