Research map / Architectures and efficiency / Pruning and efficient tuning
Task Adaptive Parameter Sharing for Multi-Task Learning
Check 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 |
|---|---|---|---|
| adaptively select task-specific layers instead of manually freezing layers at a matched parameter budget (ablation) | TAPS on ResNet-50 · ImageNet-to-Sketch | top-1 accuracy: 92.35 / 73.32 / 78.68 / 87.62 / 81.01 -> 96.68 / 76.94 / 80.74 / 89.76 / 82.65 | Helped |
| adapt only batch-normalization parameters (ablation) | ResNet-50 · ImageNet-to-Sketch | top-1 accuracy: 89.14 / 61.74 / 65.90 / 55.52 / 63.46 -> 91.07 / 70.06 / 78.47 / 79.41 / 76.25 | Helped |
| replace ImageNet pretraining with Places-365 pretraining (ablation) | TAPS · ImageNet-to-Sketch | average accuracy: drop 2.77% | Hurt |
| increase sparsity strength from λ=0.25 to λ=1.0 (ablation) | TAPS · Visual Decathlon | mean accuracy: 78.70 -> 77.77 | Hurt |
| train the task models jointly rather than incrementally (ablation) | TAPS · DomainNet | mean accuracy: 66.84 -> 66.51 | No clear effect |
| use the memory-efficient joint-training variant (ablation) | TAPS · DomainNet | mean accuracy: 66.88 -> 66.84 | No clear effect |
| add task-specific weight deltas rather than use AdaShare's select-or-skip policy | TAPS versus AdaShare · DomainNet joint MTL | mean accuracy: 62.88 -> 66.84 | Helped |
| use TAPS rather than AdaShare | DomainNet incremental MTL · DomainNet | mean accuracy: 64.33 -> 66.51 | Helped |
Closest papers by meaning
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.88)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.87)
- Quantifying Task Priority for Multi-Task Optimization (CVPR 2024 · similarity 0.87)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.87)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.87)
- MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning (CVPR 2024 · similarity 0.86)
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.