Research map / Architectures and efficiency / CNN design
Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks
Zhaoyang Zhang, Jingyu Li, Wenqi Shao, Zhanglin Peng, Ruimao Zhang, Xiaogang Wang, Ping Luo
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 group numbers per layer with DGConv instead of fixed 32 groups (b=32) | ResNeXt50 · ImageNet val | top-1 accuracy: 77.9 -> 78.4 | Improved |
| learn group numbers per layer with DGConv instead of fixed 32 groups (b=32) | ResNeXt101 · ImageNet val | top-1 accuracy: 79.1 -> 79.9 | Improved |
| increase scale constant b from 32 to 128 (less group-conv complexity) (ablation) | G-ResNeXt50 · ImageNet val | top-1 accuracy: 78.4 -> 77.8 | Hurt |
| increase scale constant b from 32 to 256 (less group-conv complexity) (ablation) | G-ResNeXt101 · ImageNet val | top-1 accuracy: 79.9 -> 79.0 | Hurt |
| reduce group conv FLOPs from 0.70x to 0.47x via b=64 (ablation) | G-ResNeXt101 · ImageNet val | top-1 accuracy: 79.9 -> 79.8 | No clear effect |
| replace uniform groups with learned DGConv groups at b=96 (0.27x FLOPs) | ResNeXt50 · ImageNet val | top-1 accuracy: 77.9 -> 78.0 | No clear effect |
| use DGConv at 2x ResNet101 complexity (b=2) instead of ResNeXt101 64x4d | ResNeXt101 · ImageNet val | top-1 accuracy: 79.8 -> 80.1 | Improved |
| re-train G-ResNeXt101 with a different random seed | G-ResNeXt101 (b=32) · ImageNet val | top-1 accuracy: 79.9 -> 79.8 | No clear effect |
| fix ResNeXt50 group numbers to those learned by DGConv (b=32) and train a plain network (ablation) | ResNeXt50 · ImageNet val | top-1 accuracy: 77.9 -> 78.3 | Improved |
| fix ResNeXt101 group numbers to those learned by DGConv (b=32) and train a plain network (ablation) | ResNeXt101 · ImageNet val | top-1 accuracy: 79.1 -> 79.8 | Improved |
Closest papers by meaning
- Dynamic Group Convolution for Accelerating Convolutional Neural Networks (ECCV 2020 · similarity 0.93)
- On the VC dimension of deep group convolutional neural networks (NeurIPS 2025 · similarity 0.88)
- Variational Partial Group Convolutions for Input-Aware Partial Equivariance of Rotations and Color-Shifts (ICML 2024 · similarity 0.87)
- Continuous Rotation Group Equivariant Network Inspired by Neural Population Coding (AAAI 2024 · similarity 0.86)
- Dynamic Group Normalization: Spatio-Temporal Adaptation to Evolving Data Statistics (CVPR 2025 · similarity 0.86)
- Group Downsampling with Equivariant Anti-aliasing (ICLR 2025 · similarity 0.85)
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.
