Research map / Architectures and efficiency / CNN design
Impact of Aliasing on Generalization in Deep Convolutional Networks
Cristina Vasconcelos, Hugo Larochelle, Vincent Dumoulin, Rob Romijnders, Nicolas Le Roux, Ross Goroshin
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 pre-filter to strided skip connections (ablation) | ResNet-50 · ImageNet val (90 epochs) | top-1 accuracy (%): 76.49 -> 77.14 | Improved |
| add post-filter to strided skip connections (ablation) | ResNet-50 · ImageNet val (90 epochs) | top-1 accuracy (%): 76.49 -> 77.15 | Improved |
| insert low-pass filters in all convolutional layers (ablation) | ResNet-50 · ImageNet val (90 epochs) | top-1 accuracy (%): 76.49 -> 61.00 | Hurt |
| anti-alias the conv1 layer that already has a 7x7 kernel (ablation) | ResNet-50 · ImageNet val (90 epochs) | top-1 accuracy (%) | Hurt |
| add post-filter anti-aliasing at all strided critical paths (ablation) | ResNet-50 · ImageNet val (90 epochs) | top-1 accuracy (%): 76.49 -> 77.47 | Improved |
| add anti-aliasing (ablation) | ResNet-50 · ImageNet val (180 epochs) | top-1 accuracy (%): 77.36 -> 77.76 | Improved |
| add random augmentation (ablation) | ResNet-50 · ImageNet val (180 epochs) | top-1 accuracy (%): 77.36 -> 77.32 | No clear effect |
| add anti-aliasing on top of random augmentation (ablation) | ResNet-50 · ImageNet val (180 epochs) | top-1 accuracy (%): 77.32 -> 78.45 | Improved |
| add anti-aliasing (k=3) (ablation) | EfficientNet-B0 · ImageNet val (3-seed mean, appendix table) | top-1 accuracy (%): 76.40 -> 76.65 | Improved |
| add anti-aliasing (k=3) with random augmentation (ablation) | EfficientNet-B0 · ImageNet val (3-seed mean, appendix table) | top-1 accuracy (%): 76.98 -> 77.17 | Improved |
| use anti-aliasing instead of Zhang's filters | ResNet-50 · ImageNet-R | top-1 accuracy (%): 24.4 -> 25.2 | Improved |
| add anti-aliasing and GELU to SUR | SUR · Meta-Dataset (all training sources) | average rank: 3.40 -> 1.65 | Improved |
Closest papers by meaning
- Understanding deep learning requires rethinking generalization (ICLR 2017 · similarity 0.89)
- Generalization bounds for deep convolutional neural networks (ICLR 2020 · similarity 0.87)
- Delving Deeper into Anti-Aliasing in ConvNets (IJCV 2022 · similarity 0.87)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · similarity 0.86)
- Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization (CVPR 2025 · similarity 0.86)
- Can Biases in ImageNet Models Explain Generalization? (CVPR 2024 · similarity 0.86)
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.
