Research map / Architectures and efficiency / Neural architecture search
Dynamic Neural Network for Multi-Task Learning Searching Across Diverse Network Topologies
Wonhyeok Choi, Sunghoon Im
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 the searched task-adaptive network with M = 7 instead of AdaShare | Multi-task learning · NYU-v2 | Δ_T: +8.9 vs +11.1 | Improved |
| Use the searched task-adaptive network with M = 9 instead of Cross-Stitch | Multi-task learning · Cityscapes | Δ_T: +2.8 vs +6.6 | Improved |
| Use the searched task-adaptive network with M = 9 instead of Cross-Stitch | Multi-task learning · PASCAL-Context | Δ_T: +0.1 vs +1.4 | Improved |
| Use the proposed network with M = 5 instead of Gumbel-Matrix | 20-way multi-task classification · Omniglot | Test Acc. (%): 93.52 -> 95.71 | Improved |
| Add flow restriction (ablation) | Ours (M = 7) · NYU-v2 | Δ_T: +11.0 -> +11.1 | No clear effect |
| Add read-in/out layers (ablation) | Ours (M = 7) · NYU-v2 | Δ_T: +10.1 -> +11.1 | Improved |
| Add flow-based reduction (ablation) | Ours (M = 7) · NYU-v2 | Δ_T: +11.5 -> +11.1 | Hurt |
| Add flow-based reduction (ablation) | Ours (M = 7) · NYU-v2 | # of Param: 6.50 -> 1.31 | Improved |
| Add squeeze loss (ablation) | Ours (M = 7) · NYU-v2 | Δ_T: +10.9 -> +11.1 | Improved |
| Add warm-up stage (ablation) | Three-stage training · NYU-v2 | Δ_T: -4.3 -> 0.0 | Improved |
| Add search and flow-based reduction stages (ablation) | Three-stage training · NYU-v2 | Δ_T: -6.1 -> 0.0 | Improved |
| Add fine-tuning stage (ablation) | Three-stage training · NYU-v2 | Δ_T: -8.9 -> 0.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AdaShare — AdaShare: Learning What To Share For Efficient Deep Multi-Task Learning (NeurIPS 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Multi-task Graph Neural Architecture Search with Task-aware Collaboration and Curriculum (NeurIPS 2023 · similarity 0.89)
- MTL-NAS: Task-Agnostic Neural Architecture Search Towards General-Purpose Multi-Task Learning (CVPR 2020 · similarity 0.88)
- Controllable Dynamic Multi-Task Architectures (CVPR 2022 · similarity 0.87)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.85)
- DMTG: One-Shot Differentiable Multi-Task Grouping (ICML 2024 · similarity 0.85)
- No More Tuning: Prioritized Multi-Task Learning with Lagrangian Differential Multiplier Methods (AAAI 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.
