Research map / Architectures and efficiency / CNN design
Interleaved Group Convolutions
Ting Zhang, Guo-Jun Qi, Bin Xiao, Jingdong Wang
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 interleaved group convolution (IGC-L24M2) instead of regular convolution W18 | CIFAR-100 depth 38 · CIFAR-100 | accuracy (%): 65.00 -> 69.56 | Improved |
| use IGC-L4M8 instead of RegConv-W18 | CIFAR-100 depth 38 · CIFAR-100 | accuracy (%): 65.00 -> 67.33 | Improved |
| use IGC-L24M2 instead of RegConv-W18 | CIFAR-10 depth 20 · CIFAR-10 | accuracy (%): 92.55 -> 92.84 | Improved |
| use IGC-L4M8 instead of RegConv-W18 | CIFAR-10 depth 8 · CIFAR-10 | accuracy (%): 90.30 -> 89.89 | Mixed |
| replace summation fusion with interleaved group convolution (ablation) | CIFAR-100 depth 20 · CIFAR-100 | accuracy (%): 59.33 -> 69.45 | Improved |
| increase primary partitions to L=40, M=2 from Xception-like L=64, M=1 (ablation) | IGC 8-layer net · CIFAR-100 | accuracy (%): 63.07 -> 63.89 | Improved |
| use single primary partition L=1 (regular convolution) with M=23 (ablation) | IGC 8-layer net · CIFAR-100 | accuracy (%): 51.11 vs 63.89 at L=40 M=2 | Hurt |
| add identity mappings to IGC-L24M2 | IGC-L24M2 depth 98 · CIFAR-100 | accuracy (%): 76.15 vs RegConv-W18+Ident. 75.30 | Improved |
| use IGC-L16M16+Ident. instead of ResNet C=69 | ResNet-18 · ImageNet val | top-1 error (%): 30.58 -> 29.40 | Improved |
| use IGC-L100M2+Ident. with fewer parameters | ResNet-18 · ImageNet val | top-1 error (%): 31.06 -> 26.95 | Improved |
| add a nonlinear activation between primary and secondary group convolutions (ablation) | IGC block · classification | accuracy | Hurt |
| swap kernel sizes (1x1 primary, 3x3 secondary) (ablation) | IGC block · classification | accuracy | No clear effect |
Closest papers by meaning
- Interleaved Structured Sparse Convolutional Neural Networks (CVPR 2018 · similarity 0.90)
- Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks (ICCV 2019 · similarity 0.90)
- Dynamic Group Convolution for Accelerating Convolutional Neural Networks (ECCV 2020 · similarity 0.89)
- Exploiting Redundancy: Separable Group Convolutional Networks on Lie Groups (ICML 2022 · similarity 0.89)
- Group Downsampling with Equivariant Anti-aliasing (ICLR 2025 · similarity 0.87)
- Unitary Convolutions for Learning on Graphs and Groups (NeurIPS 2024 · 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.
