Research map / Architectures and efficiency / CNN design
Selective Kernel Networks
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 ResNeXt bottleneck convolutions with SK convolutions using adaptive kernel selection | ResNeXt-50 · ImageNet validation, 224×224 | top-1 error (%): 22.23 -> 20.79 | Helped |
| Replace ResNeXt bottleneck convolutions with SK convolutions using adaptive kernel selection | ResNeXt-101 · ImageNet validation, 224×224 | top-1 error (%): 21.11 -> 20.19 | Helped |
| Replace 3×3 depthwise convolutions with SK units | ShuffleNetV2 · ImageNet validation, single 224×224 crop | top-1 error (%): 38.41 -> 35.35; 30.57 -> 28.36 | Helped |
| Use SK convolution in place of the ResNeXt grouped convolution | ResNeXt-29, 16×32d · CIFAR-10 test | top-1 error (%): 3.87 -> 3.47 | Helped |
| Use SK convolution in place of the ResNeXt grouped convolution | ResNeXt-29, 16×32d · CIFAR-100 test | top-1 error (%): 18.56 -> 17.33 | Helped |
| Use SK attention instead of simple summation across different kernel paths (ablation) | SKNet-50 with K3 and K5 · ImageNet validation | top-1 error (%): 21.76 -> 20.79 | Helped |
| Add a second, different-sized kernel path (ablation) | K3-only model · ImageNet validation | top-1 error (%): 22.23 -> 21.76 | Helped |
| Add SK attention to K3 and K7 paths (ablation) | SKNet-50 · ImageNet validation | top-1 error (%): 21.82 -> 20.97 | Helped |
| Add SK attention to K5 and K7 paths (ablation) | SKNet-50 · ImageNet validation | top-1 error (%): 23.64 -> 23.09 | Helped |
| Add a third K7 path to an SK model using K3 and K5 (ablation) | SKNet-50 · ImageNet validation | top-1 error (%): 20.79 -> 20.76 | Helped |
| Use a 7×7 second-branch kernel instead of the selected 5×5 effective kernel (ablation) | SKNet-50 · ImageNet validation | top-1 error (%): 20.79 -> 21.18 | Hurt |
| Use dilation 3 rather than dilation 2 in the second 3×3 branch (ablation) | SKNet-50 · ImageNet validation | top-1 error (%): 20.79 -> 20.97 | Hurt |
Closest papers by meaning
- Receptive Fields As Experts in Convolutional Neural Architectures (ICML 2024 · similarity 0.87)
- Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs (CVPR 2022 · similarity 0.87)
- Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.86)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.85)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.85)
- From Kernels to Features: A Multi-Scale Adaptive Theory of Feature Learning (ICML 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.