Research map / Architectures and efficiency / CNN design
Design of Kernels in Convolutional Neural Networks for Image Classification
Zhun Sun, Mete Ozay, Takayuki Okatani
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 quasi-hexagonal kernels instead of 3x3 square | BASE (ImageNet CNN) · ILSVRC-2012 val | top-1 error (%): 31.2 -> 29.2 | Improved |
| use quasi-hexagonal kernels instead of 3x3 square | BASE (ImageNet CNN) · ILSVRC-2012 val | top-5 error (%): 12.3 -> 11.1 | Improved |
| use quasi-hexagonal kernels instead of 3x3 square (fewer params) | BASE-A · CIFAR-10 | test error (%): 9.02 -> 9.10 | No clear effect |
| use random zeroed weights in 3x3 kernels (fragmented kernel) instead of quasi-hexagonal (ablation) | QH-A · CIFAR-10 | test error (%): 9.10 -> 9.89 | Hurt |
| use random kernel shape patterns per layer instead of one fixed pattern (ablation) | QH-A · CIFAR-10 | test error (%): 9.40 -> 9.10 | Improved |
| add extra dropout before global average pooling (ablation) | BASE-A · CIFAR-10 | test error (%): 9.02 -> 8.71 | Improved |
| add extra dropout before global average pooling (ablation) | QH-A · CIFAR-10 | test error (%): 9.10 -> 8.79 | Improved |
| use larger max-pool kernel (pool4) (ablation) | QH-A · CIFAR-10 | test error (%): 9.10 -> 9.00 | Improved |
| use larger max-pool kernel (pool4) (ablation) | BASE-A · CIFAR-10 | test error (%): 9.02 -> 8.87 | Improved |
| widen QH network to match BASE-A parameter count (QH-B) and larger QH-C | BASE-A-AD · CIFAR-100 | test error (%): 31.2 -> 29.77 | Improved |
| use QH kernels with more feature maps (QH-C-AD) | BASE-A-AD · CIFAR-10 | test error (%): 8.71 -> 8.42 | Improved |
| use QH-B-AD instead of base when only 1K training samples | BASE-AD · CIFAR-10 (1K training samples) | test error (%): 44.9 -> 41.8 | Improved |
Closest papers by meaning
- End-to-End Kernel Learning with Supervised Convolutional Kernel Networks (NeurIPS 2016 · similarity 0.89)
- Shape or Texture: Understanding Discriminative Features in CNNs (ICLR 2021 · similarity 0.86)
- Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs (CVPR 2022 · similarity 0.86)
- Structured Receptive Fields in CNNs (CVPR 2016 · similarity 0.86)
- Convolution of Convolution: Let Kernels Spatially Collaborate (CVPR 2022 · similarity 0.86)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · 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.
