Research map / Architectures and efficiency / CNN design
Spectral Leakage and Rethinking the Kernel Size in CNNs
Nergis Tomen, Jan C. van Gemert
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 |
|---|---|---|---|
| add Hamming window to 7x7 kernels | ResNet18 · ImageNet val (top-1) | accuracy (%): 70.30 -> 71.54 | Improved |
| add Hamming window to 7x7 kernels | ResNet34 · ImageNet val (top-1) | accuracy (%): 72.20 -> 74.54 | Improved |
| add Hamming window to 7x7 kernels | ResNet50 · ImageNet val (top-1) | accuracy (%): 75.77 -> 76.80 | Improved |
| add Hamming window to 7x7 kernels | VGG-16 · ImageNet val (top-1) | accuracy (%): 73.47 -> 75.11 | Improved |
| enlarge kernels from default to 7x7 without windowing | ResNet34 · ImageNet val (top-1) | accuracy (%): 73.12 -> 72.20 | Hurt |
| enlarge kernels from default to 7x7 without windowing | VGG-13 · ImageNet val (top-1) | accuracy (%): 72.39 -> 72.56 | No clear effect |
| add Hamming window with 7x7 kernels | CNN baseline · CIFAR-10 under DeepFool | accuracy (%): 20.06 -> 32.64 | Improved |
| add Hamming window with 5x5 kernels | CNN baseline · CIFAR-10 under DeepFool | accuracy (%): 24.85 -> 23.20 | Hurt |
| add Hamming window with 7x7 kernels | CNN baseline · CIFAR-10 clean | accuracy (%): 83.36 -> 86.38 | Improved |
| add Hamming window with 7x7 kernels | CNN baseline · CIFAR-10 spatial attack (tr 12.5, rot 22.5) | accuracy (%): 44.59 -> 53.03 | Improved |
| add Hamming window with 9x9 kernels | CNN baseline · CIFAR-10 spatial attack (tr 25.0, rot 45.0) | accuracy (%): 18.13 -> 24.55 | Improved |
| add Hamming window to the first layer only (ablation) | CNN baseline · CIFAR-10 validation accuracy | accuracy (%) | Improved |
Closest papers by meaning
- Convolutional Spectral Kernel Learning with Generalization Guarantees (Abstract Reprint) (AAAI 2024 · similarity 0.86)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.85)
- A Closer Look at Fourier Spectrum Discrepancies for CNN-Generated Images Detection (CVPR 2021 · similarity 0.85)
- FrequencyLowCut Pooling – Plug & Play against Catastrophic Overfitting (ECCV 2022 · similarity 0.84)
- KernelDNA: Cross-Layer Kernel Sharing via Decoupled Neural Adapters (NeurIPS 2026 · similarity 0.84)
- Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness (ICML 2024 · similarity 0.83)
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.
