Research map / Architectures and efficiency / RNNs and normalization
A Modulation Module for Multi-task Learning with Applications in Image Retrieval
Xiangyun Zhao, Haoxiang Li, Xiaohui Shen, Xiaodan Liang, Ying 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 |
|---|---|---|---|
| add task-specific modulation modules to a fully shared network | fully shared network · CelebA retrieval (20 attributes) | average triplet accuracy: 69.4% -> 84.86% | Improved |
| add modulation modules instead of independent task-specific branches (IB-256) | branch baseline · CelebA retrieval (20 attributes) | average triplet accuracy: 83.69% -> 84.86% | Improved |
| add modulation modules compared with independent small branches (IB-25), same complexity | IB-25 · CelebA retrieval (40 attributes) | average triplet accuracy: 78.22% -> 85.75% | Improved |
| train shared parameters jointly with modulation instead of fixing IB-256 shared parameters and training only modulation (ablation) | Modulation module · CelebA retrieval (20 attributes) | average triplet accuracy: 76.32% -> 84.86% | Improved |
| modulate more layers (from block5 to block3) (ablation) | Modulation module · CelebA retrieval (7 attributes) | average triplet accuracy: 81.48% -> 82.98% | Improved |
| extend modulation from block3 to block2 (ablation) | Modulation module · CelebA retrieval (7 attributes) | average triplet accuracy: 82.98% -> 82.62% | No clear effect |
| use channel-wise projection matrix instead of scaling vector (ablation) | Modulation module (block4) · CelebA retrieval (7 attributes) | average triplet accuracy: 82.75% -> 82.79% | No clear effect |
| add modulation to the joint smile+young+open mouth network | fully shared 3-task network · CelebA toy (smile) | smile accuracy: 84.71% -> 94.03% | Improved |
| add relevance regularization loss on modulation weights (ablation) | Modulation module · CelebA toy (smile) | smile accuracy: 94.03% -> 94.94% | No clear effect |
| add modulation to joint smile+young training | joint smile+young network · CelebA toy (smile/young pair) | update compliance ratio: 22.1% -> 50.63% | Improved |
| add modulation to joint smile+open mouth training | joint smile+open mouth network · CelebA toy (smile/open-mouth pair) | update compliance ratio: 43.71% -> 52.77% | Improved |
| add modulation to the shared network over 4 product tasks | CSN · UT-Zappos50K retrieval (4 tasks) | average triplet accuracy: 86.73% -> 88.17% | Improved |
Closest papers by meaning
- Cross-Stitch Networks for Multi-Task Learning (CVPR 2016 · similarity 0.87)
- Toward Unified Expertise: Learning a Single Vision Model From Diverse Perception (TPAMI 2025 · similarity 0.87)
- Multi-Task Learning as Multi-Objective Optimization (NeurIPS 2018 · similarity 0.87)
- Mod-Squad: Designing Mixtures of Experts As Modular Multi-Task Learners (CVPR 2023 · similarity 0.87)
- MmAP: Multi-Modal Alignment Prompt for Cross-Domain Multi-Task Learning (AAAI 2024 · similarity 0.86)
- Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · 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.
