Research map / Architectures and efficiency / CNN design
Dynamic Capacity Networks
Read the paperCheck 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 |
|---|---|---|---|
| Add hints objective aligning coarse and fine features (ablation) | DCN · Cluttered MNIST | test error (): 1.71 -> 1.39 | Helped |
| Use DCN attention on 8 patches instead of fine model on full input | DCN · Cluttered MNIST | test error (): 1.70 -> 1.39 | Helped |
| Use DCN instead of fine model | DCN · Cluttered MNIST | multiplications (M): 84.5M -> 27.7M | Helped |
| Entropy-weighted soft attention over coarse features | coarse model · SVHN | test error (): 40.6 -> 31.4 | Helped |
| Use DCN with 6 fine patches instead of fine model | DCN · SVHN | test error (): 25.2 -> 20.0 | Helped |
| Increase scales from 1 to 2 for DCN | DCN · SVHN | test error (): 20.0 -> 18.2 | Helped |
| Increase to 9 patches and 3 scales | DCN · SVHN | test error (): 18.2 -> 16.6 | Helped |
| Take all possible patches at 3 scales | DCN · SVHN | test error (): 16.6 -> 11.6 | Helped |
| Increase scales from 1 to 3 for coarse model | coarse model · SVHN | test error (): 40.6 -> 40.0 | No clear effect |
| Increase scales from 1 to 3 for fine model | fine model · SVHN | test error (): 25.2 -> 23.3 | Helped |
| Increase number of patches beyond about 10 | DCN · Cluttered MNIST | test error | No clear effect |
| Use DCN instead of fine model for inference time | DCN · SVHN largest 100 images | inference time (ms): 62.6 -> 10.8 | Helped |
Closest papers by meaning
- Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.87)
- Dynamic Deep Neural Networks: Optimizing Accuracy-Efficiency Trade-Offs by Selective Execution (AAAI 2018 · similarity 0.87)
- Dynamic Filter Networks (NeurIPS 2016 · similarity 0.86)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.84)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.84)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 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.