Research map / Architectures and efficiency / CNN design
Stochastic Filter Groups for Multi-Task CNNs: Learning Specialist and Generalist Convolution Kernels
Felix J.S. Bragman, Ryutaro Tanno, Sebastien Ourselin, Daniel C. Alexander, Jorge Cardoso
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 |
|---|---|---|---|
| learn kernel allocation (SFG) instead of fixed 1/3 mask (ablation) | VGG-11 multi-task · UTKFace full data | age MAE: 7.67 -> 6.00 | Improved |
| learn kernel allocation (SFG) instead of fixed 1/3 mask (ablation) | VGG-11 multi-task · UTKFace full data | gender accuracy (%): 89.41 -> 92.46 | Improved |
| learn grouping probabilities instead of constant p=[1/3,1/3,1/3] (ablation) | VGG-11 multi-task · UTKFace full data | age MAE: 6.34 -> 6.00 | Improved |
| learn grouping probabilities instead of constant p (ablation) | VGG-11 multi-task · UTKFace small data | age MAE: 9.15 -> 8.54 | Improved |
| learn grouping probabilities instead of constant p (ablation) | VGG-11 multi-task · UTKFace small data | gender accuracy (%): 86.01 -> 87.01 | Improved |
| use SFG instead of hard parameter sharing (ablation) | VGG-11 multi-task · UTKFace full data | age MAE: 7.92 -> 6.00 | Improved |
| use SFG instead of cross-stitch | VGG-11 multi-task · UTKFace small data | gender accuracy (%): 83.72 -> 87.01 | Improved |
| use SFG instead of cross-stitch | HighResNet multi-task · prostate CT synthesis overall | PSNR: 27.86 -> 27.74 | No clear effect |
| use SFG instead of cross-stitch | HighResNet multi-task · prostate segmentation overall | DICE: 0.854 -> 0.852 | No clear effect |
| learn grouping (SFG) instead of constant mask (ablation) | HighResNet multi-task · prostate segmentation overall | DICE: 0.774 -> 0.852 | Improved |
| use hard parameter sharing multi-task instead of single-task | HighResNet · prostate segmentation overall | DICE: 0.848 -> 0.829 | Hurt |
| use SFG instead of cross-stitch on bone regions | HighResNet multi-task · prostate CT synthesis bones | PSNR: 32.27 -> 32.29 | No clear effect |
Closest papers by meaning
- Efficiently Identifying Task Groupings for Multi-Task Learning (NeurIPS 2021 · similarity 0.88)
- Mod-Squad: Designing Mixtures of Experts As Modular Multi-Task Learners (CVPR 2023 · similarity 0.88)
- DMTG: One-Shot Differentiable Multi-Task Grouping (ICML 2024 · similarity 0.88)
- Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.87)
- Multi-Task Dense Prediction via Mixture of Low-Rank Experts (CVPR 2024 · similarity 0.86)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 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.
