Research map / Architectures and efficiency / CNN design
Making Convolutional Networks Shift-Invariant Again
Read the paperCheck 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 |
|---|---|---|---|
| Add low-pass filtering before downsampling | MaxBlurDown classifier · Image classification with spatial shifts | classification consistency across spatial positions | Helped |
| Use a larger binomial blur filter (ablation) | MaxBlurDown classifier · Image classification with spatial shifts | classification consistency across spatial positions | Helped |
| Add a low-pass filter | Convolutional classifier · Classification variation across image shifts | variation in probability of correct classification | Helped |
| Increase the low-pass filter size (ablation) | Convolutional classifier · Classification variation across image shifts | variation in probability of correct classification | Helped |
| Add low-pass filtering | Convolutional classifier · Classification variation across image shifts, with and without data augmentation | variation in probability of correct classification | Helped |
| Increase the binomial filter size (ablation) | VGG13 · CIFAR10 shift-based adversary | accuracy under adversarial shifts | Helped |
| Use Bin-7 without data augmentation instead of the baseline with data augmentation (ablation) | VGG13 · CIFAR10 shift-based adversary | accuracy under adversarial shifts | Helped |
| Increase the binomial filter size (ablation) | VGG13 trained with data augmentation · CIFAR10 shift-based adversary | accuracy under adversarial shifts | Helped |
| Train with MaxBlurPool rather than MaxPool | Convolutional classifier · Learned convolutional filters | normalized total variation | Helped |
| Use more aggressive blur kernels (ablation) | Convolutional classifier · Learned convolutional filters | normalized total variation | Helped |
| Use Bin-4 or Bin-5 filters (ablation) | Labels-to-Facades image-to-image translation model · Facades test set | generated image quality | Hurt |
Closest papers by meaning
- Alias-Free Convnets: Fractional Shift Invariance via Polynomial Activations (CVPR 2023 · similarity 0.90)
- Truly Shift-Invariant Convolutional Neural Networks (CVPR 2021 · similarity 0.90)
- Alias-Free ViT: Fractional Shift Invariance via Linear Attention (NeurIPS 2025 · similarity 0.87)
- Group Downsampling with Equivariant Anti-aliasing (ICLR 2025 · similarity 0.87)
- Learning Strides in Convolutional Neural Networks (ICLR 2022 · similarity 0.86)
- Don’t Think It Twice: Exploit Shift Invariance for Efficient Online Streaming Inference of CNNs (AAAI 2025 · 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.