Research map / Architectures and efficiency / CNN design
Integrating Multiple Receptive Fields Through Grouped Active Convolution
Yunho Jeon, Junmo Kim
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 3x3 conv with ACU | ResNet-29 bottleneck · CIFAR-10 | test error (%): 6.16 -> 5.56 | Improved |
| enlarge receptive field via dilation 2 in last stage (ablation) | ResNet-29 baseline · CIFAR-10 | test error (%): 6.16 -> 6.28 | Hurt |
| fix ACU shape to learned shape and retrain weights (ablation) | ACU network · CIFAR-10 | test error (%): 5.56 -> 5.44 | Improved |
| replace grouped conv with grouped ACU | 16x4d ResNet · CIFAR-10 | test error (%): 5.38 -> 4.93 | Improved |
| increase cardinality (4d to 1d) with naive 3x3 conv (ablation) | grouped conv network · CIFAR-10 | test error (%): 5.38 -> 5.75 | Hurt |
| increase cardinality (4d to 1d) with ACU (ablation) | grouped ACU network · CIFAR-10 | test error (%): 4.93 -> 4.9 | No clear effect |
| add warming-up for shared-position ACU (ablation) | ResNet/ACU · CIFAR-10 | test error (%): 5.64 -> 5.56 | Improved |
| add warming-up for depthwise ACU (ablation) | x1d/ACU · CIFAR-10 | test error (%): 4.9 -> 5.04 | Hurt |
| use multiple position sets instead of one shared (ablation) | depthwise ACU x1d · CIFAR-10 | test error (%): 5.2 -> 4.9 | Improved |
| replace Inception blocks with 1x1 conv plus depthwise ACU block | Inception+BN · ImageNet | top-1 accuracy (%): 72.0 -> 72.1 | Improved |
| replace last depthwise conv per block with depthwise ACU | Xception · ImageNet | top-1 accuracy (%): 75.6 -> 75.9 | Improved |
| replace depthwise conv with depthwise ACU | MobileNet v2 x1.0 · ImageNet | top-1 accuracy (%): 71.3 -> 72.1 | Improved |
Closest papers by meaning
- Active Convolution: Learning the Shape of Convolution for Image Classification (CVPR 2017 · similarity 0.91)
- Understanding the Effective Receptive Field in Deep Convolutional Neural Networks (NeurIPS 2016 · similarity 0.86)
- Exploiting Redundancy: Separable Group Convolutional Networks on Lie Groups (ICML 2022 · similarity 0.86)
- Spatially-Adaptive Filter Units for Deep Neural Networks (CVPR 2018 · similarity 0.86)
- Group-Wise Deep Object Co-Segmentation With Co-Attention Recurrent Neural Network (ICCV 2019 · similarity 0.86)
- Receptive Fields As Experts in Convolutional Neural Architectures (ICML 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.
