Research map / Architectures and efficiency / CNN design
Dynamic Dual Gating Neural Networks
Fanrong Li, Gang Li, Xiangyu He, Jian Cheng
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 both spatial and channel gating instead of only one (ablation) | DGNet · CIFAR-10 (ResNet-20/32) | top-1 accuracy | Improved |
| use spatial gating only instead of channel gating only (ablation) | DGNet · CIFAR-10 (ResNet-20/32) | top-1 accuracy | Mixed |
| use 1x1 fine-grained spatial tiles instead of 2x2 or hybrid tiles (ablation) | DGNet · CIFAR-10 (ResNet-20/32) | top-1 accuracy | Hurt |
| use 8x8 tiles in all layers instead of hybrid tile sizes (ablation) | DGNet · CIFAR-10 (ResNet-20/32) | top-1 accuracy | Hurt |
| use hybrid tile sizes (per-stage) instead of uniform tile size (ablation) | DGNet · CIFAR-10 (ResNet-20/32) | top-1 accuracy | Improved |
| avoid applying spatial and channel gating in the same layer (proposed integration) instead of inserting spatial gating in all convolutions (ablation) | DGNet · CIFAR-10 (ResNet-20/32) | top-1 accuracy | Improved |
| use a shared tiled mask (t=4) instead of a fine-grained mask (t=1) when dilating for 3x3 convolutions (ablation) | spatial gating mask · 28x28 mask visualization | mask sparsity after dilation: 47.84% -> 0.26% for t=1; 44.90% -> 23.85% for t=4 | Improved |
| apply dual gating with 50% target to ResNet-18 | ResNet-18 · ImageNet val | top-1 accuracy: 69.76 -> 70.12 | Improved |
| apply dual gating with 60% target to ResNet-50 | ResNet-50 · ImageNet val | top-1 accuracy: 76.13 -> 76.41 | Improved |
| raise target sparsity from 60% to 70% | DGNet ResNet-34 · ImageNet val | top-1 accuracy: 73.01 -> 71.95 | Hurt |
| apply dual gating (60%) to RetinaNet backbone | RetinaNet · COCO 2017 | AP0.5:0.95: 36.5 -> 36.1 | No clear effect |
| apply dual gating (60%) to Faster-RCNN backbone | Faster-RCNN FPN · COCO 2017 | AP0.5:0.95: 37.4 -> 37.2 | No clear effect |
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.
- HRank — HRank: Filter Pruning Using High-Rank Feature Map (CVPR 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
- Channel Gating Neural Networks (NeurIPS 2019 · similarity 0.88)
- Precision Gating: Improving Neural Network Efficiency with Dynamic Dual-Precision Activations (ICLR 2020 · similarity 0.87)
- Dynamic Neural Networks: A Survey (TPAMI 2021 · similarity 0.86)
- Is Kernel Prediction More Powerful than Gating in Convolutional Neural Networks? (ICML 2024 · similarity 0.84)
- Jointly-Learned Exit and Inference for a Dynamic Neural Network (ICLR 2024 · similarity 0.84)
- Rethinking Neural Nonlinearity as Gating (NeurIPS 2026 · 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.
