Research map / Label-efficient and robust learning / Recognition and long-tail
Mitigating Negative Transfer in Multi-Task Learning with Exponential Moving Average Loss Weighting Strategies (Student Abstract)
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 |
|---|---|---|---|
| Scale each task loss by the reciprocal of its EMA | Multi-task learning · CelebA | mean F1: 0.772 -> 0.782 | Helped |
| Combine EMA loss scaling with training-rate weighting (DWEMA) | Multi-task learning · CelebA | mean F1: 0.772 -> 0.788 | Helped |
| Combine EMA loss scaling with training-rate weighting (REMA) | Multi-task learning · AffWild2 | overall performance: 0.760 -> 0.764 | No clear effect |
| Use EMA loss scaling without training-rate weighting | Multi-task learning · AffWild2 | overall performance: 0.760 -> 0.800 | Helped |
| Use EMA loss scaling without training-rate weighting | Multi-task learning · AffectNet | overall performance: 0.813 -> 0.898 | Helped |
| Set EMA β to 0.1 rather than 0.5 (ablation) | EMA loss weighting · AffWild2 | overall performance: 0.757 -> 0.800 | Helped |
| Set EMA β to 0.1 rather than 0.5 (ablation) | EMA loss weighting · AffectNet | overall performance: 0.839 -> 0.898 | Helped |
| Set EMA β to 0.2 rather than 0.5 (ablation) | EMA loss weighting · CelebA | overall performance: 0.754 -> 0.783 | Helped |
| Set DWA temperature T to 0.5 rather than 2 (ablation) | DWA · AffWild2 | overall performance: 0.741 -> 0.76 | Helped |
| Set DWA temperature T to 1 (ablation) | DWA · AffectNet | overall performance: 0.813 -> 0.081 | Hurt |
| Add training-rate weighting to EMA loss scaling | EMA loss weighting · AffWild2 | overall performance: 0.800 -> 0.775 (DWEMA); 0.800 -> 0.764 (REMA) | Hurt |
| Add training-rate weighting to EMA loss scaling | EMA loss weighting · AffectNet | overall performance: 0.898 -> 0.868 (DWEMA); 0.898 -> 0.869 (REMA) | Hurt |
Closest papers by meaning
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.87)
- Robust Multi-Task Learning with Excess Risks (ICML 2024 · similarity 0.86)
- Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining (NeurIPS 2025 · similarity 0.85)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.85)
- Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction (CVPR 2025 · similarity 0.85)
- MTL-UE: Learning to Learn Nothing for Multi-Task Learning (ICML 2025 · similarity 0.84)
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.