Research map / Architectures and efficiency / CNN design
Differentiable Dynamic Wirings for Neural Networks
Kun Yuan, Quanquan Li, Shaopeng Guo, Dapeng Chen, Aojun Zhou, Fengwei Yu, Ziwei Liu
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 |
|---|---|---|---|
| replace static wiring with dynamic instance-aware wiring | MobileNetV2-1.0 · ImageNet val | top-1 accuracy (%): 72.60 -> 73.54 | Improved |
| replace static wiring with dynamic instance-aware wiring | ResNet101 · ImageNet val | top-1 accuracy (%): 78.29 -> 79.90 | Improved |
| replace static wiring with dynamic instance-aware wiring | RegNet-X-600M · ImageNet val | top-1 accuracy (%): 74.03 -> 74.68 | Improved |
| use dynamic wiring on deeper ResNets (larger search space) | ResNet18 to ResNet101 · ImageNet val | top-1 accuracy (%): 1.02% to 1.61% | Improved |
| replace static wiring with dynamic wiring in detector backbone | ResNet101 backbone, Faster R-CNN FPN · COCO minival | box AP (%): 38.59 -> 41.32 | Improved |
| replace static wiring with dynamic wiring in detector backbone | ResNet50 backbone, Faster R-CNN FPN · COCO minival | box AP (%): 36.42 -> 38.12 | Improved |
| use dynamic DDW instead of RandWire random static graphs (WS) | MobileNetV2-size networks · ImageNet val (small computation regime) | top-1 accuracy (%): 72.26 -> 73.52 | Improved |
| use DDW instead of InstaNAS-ImgNet-A | MobileNetV2 · ImageNet val | top-1 accuracy (%): 71.9 -> 73.5 | Improved |
| add learnable static edge weights (alpha) (ablation) | ResNet-50 · ImageNet val | top-1 accuracy (%): 76.70 -> 77.00 | Improved |
| add dynamic routing on top of static learnable edge weights (ablation) | ResNet-50 · ImageNet val | top-1 accuracy (%): 77.00 -> 78.28 | Improved |
| add dynamic routing on top of static learnable edge weights (ablation) | ResNet-101 · ImageNet val | top-1 accuracy (%): 78.64 -> 79.90 | Improved |
| add dynamic routing on top of static learnable edge weights (ablation) | MobileNetV2-1.0 · ImageNet val | top-1 accuracy (%): 72.86 -> 73.54 | 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.
- DARTS — DARTS-: Robustly Stepping out of Performance Collapse Without Indicators (ICLR 2021) · 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
- Discovering Neural Wirings (NeurIPS 2019 · similarity 0.88)
- Exploring Randomly Wired Neural Networks for Image Recognition (ICCV 2019 · similarity 0.87)
- MaskConnect: Connectivity Learning by Gradient Descent (ECCV 2018 · similarity 0.87)
- Deciding How to Decide: Dynamic Routing in Artificial Neural Networks (ICML 2017 · similarity 0.86)
- Adaptive Activation Thresholding: Dynamic Routing Type Behavior for Interpretability in Convolutional Neural Networks (ICCV 2019 · similarity 0.86)
- Input Space Mode Connectivity in Deep Neural Networks (ICLR 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.
