Research map / Architectures and efficiency / RNNs and normalization
Deep Multitask Learning with Progressive Parameter Sharing
Haosen Shi, Shen Ren, Tianwei Zhang, Sinno Jialin Pan
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 |
|---|---|---|---|
| apply progressive parameter-sharing masks (MPPS) on top of hard-parameter sharing | HPS · NYUv2 | Delta: +4.8% -> +6.5% (Gaussian), +5.7% (Bernoulli) | Improved |
| apply MPPS on Cityscapes | HPS · Cityscapes | Delta: +0.7% -> +1.7% | Improved |
| apply MPPS on Multi-CIFAR100 | HPS (VGG16) · Multi-CIFAR100 | average accuracy: 59.23% -> 67.62% | Improved |
| apply MPPS to all layers instead of only last 6 (ablation) | MPPS · NYUv2 | Delta: +5.7% -> -11.4% | Hurt |
| restrict MPPS to last 6 layers rather than 12 or 18 (ablation) | MPPS · NYUv2 | Delta: +4.5% (12 layers), +4.2% (18 layers) -> +5.7% (6 layers) | Improved |
| add target distribution updating (ablation) | MPPS · NYUv2 | Delta: +4.8% -> +5.7% | Improved |
| progressive EC scheduling versus fixed highest exclusive capacity (Separ) (ablation) | MPPS · NYUv2 | improvement over baseline | Improved |
| add moderate frozen epochs (15, 25) to Linear scheduler (ablation) | MPPS · NYUv2 | improvement | Mixed |
| combine MPPS with existing MTL methods | DWA, Auto-lambda, Graddrop, PCGrad, Cagrad, HPS, MTAN · NYUv2 | Delta: 6/9 methods improve, e.g. MTAN 10.3 -> 10.9; Random 7.2 -> 5.5; Uncert 4.7 -> 4.3 | Mixed |
| apply MPPS to smaller/larger backbones (ResNet-18, ResNet-34) | HPS · NYUv2 | Delta: +2.7% -> +4.2% (ResNet-18); +5.7% -> +6.9% (ResNet-34) | Improved |
| MPPS overhead on HPS training | HPS · NYUv2, ResNet-50 | training time per epoch (s): 28.69 -> 39.34 | Hurt |
Closest papers by meaning
- Efficient Computation Sharing for Multi-Task Visual Scene Understanding (ICCV 2023 · similarity 0.87)
- Learning Sparse Sharing Architectures for Multiple Tasks (AAAI 2020 · similarity 0.87)
- AdaShare: Learning What To Share For Efficient Deep Multi-Task Learning (NeurIPS 2020 · similarity 0.87)
- Modular Universal Reparameterization: Deep Multi-task Learning Across Diverse Domains (NeurIPS 2019 · similarity 0.86)
- Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.85)
- Quantifying Task Priority for Multi-Task Optimization (CVPR 2024 · 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.
