Research map / Adversarial robustness and security / Certified and empirical robustness
High-Frequency Component Helps Explain the Generalization of Convolutional Neural 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 |
|---|---|---|---|
| Smooth first-layer kernels with ρ=1.0 (ablation) | Vanilla ResNet-18 · CIFAR10 test, clean and PGD ε=0.09 | accuracy: 0.856 -> 0.631; 0.002 -> 0.078 | Mixed |
| Smooth first-layer kernels with ρ=1.0 (ablation) | PGD-trained ResNet-18 · CIFAR10 test, clean and PGD ε=0.09 | accuracy: 0.707 -> 0.638; 0.038 -> 0.131 | Mixed |
| Train on LFC rather than HFC | ResNet-18 · CIFAR10 test, frequency components r=4, 8, 12, 16 | accuracy: 0.2002 -> 0.6167; 0.092 -> 0.7154; 0.0997 -> 0.7516; 0.1281 -> 0.7714 | Helped |
| Use LFC instead of original images | RetinaNet (ResNet50 + FPN) · COCO detection val, r=128 | mAP: 35.6% -> 27.5% | Hurt |
| Use HFC instead of original images | RetinaNet (ResNet50 + FPN) · COCO detection val, r=128 | mAP: 35.6% -> 10.7% | Hurt |
| Use Mix-up training (ablation) | ResNet-18 · CIFAR10 | prediction accuracy | No clear effect |
| Use Mix-up training (ablation) | ResNet-18 · CIFAR10, HFC response | accuracy on HFC | Helped |
| Use dropout on fully connected layers (ablation) | ResNet-18 · CIFAR10 | prediction accuracy | No clear effect |
| Use adversarial training (ablation) | ResNet-18 · CIFAR10 | clean prediction accuracy | Hurt |
| Use BatchNorm (ablation) | ResNet-18 · CIFAR10, LFC and HFC | component accuracy | Helped |
| Train with larger batch size (ablation) | ResNet-18 · CIFAR10 | generalization gap | Helped |
Closest papers by meaning
- Spectral Bias in Practice: The Role of Function Frequency in Generalization (NeurIPS 2022 · similarity 0.88)
- On the Frequency Bias of Generative Models (NeurIPS 2021 · similarity 0.88)
- Which Frequencies do CNNs Need? Emergent Bottleneck Structure in Feature Learning (ICML 2024 · similarity 0.87)
- Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization (CVPR 2025 · similarity 0.86)
- Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks (AAAI 2024 · similarity 0.85)
- From Linear to Nonlinear: Provable Weak-to-Strong Generalization through Feature Learning (NeurIPS 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.