Research map / Architectures and efficiency / CNN design
Why do deep convolutional networks generalize so poorly to small image transformations?
Aharon Azulay, Yair Weiss
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 |
|---|---|---|---|
| embed image at a size far from the training size (about 100 instead of about 220) (ablation) | VGG16, ResNet50, InceptionResNetV2 · ImageNet embedding protocol | P(top-1 change): approximately 10% -> approximately 30% | Hurt |
| use an atypical image with low similarity to training neighbours (ablation) | ImageNet CNNs · ImageNet crop protocol | P(top-1 change): approaches 0.3 | Hurt |
| add bin-5 antialiasing (blur before stride) | modern CNNs · ImageNet embedding protocol | P(top-1 change): about 15% | No clear effect |
| add bin-5 antialiasing, atypical images | modern CNNs · ImageNet crop protocol | P(top-1 change): up to 20% | No clear effect |
| train a readout from deeper layers instead of early layers (ablation) | pretrained CNNs · ImageNet crop protocol | P(top-1 change): below 5% -> higher with depth | Hurt |
| remove subsampling from the CNN (ablation) | AlexNet-style CNN · CIFAR-10 | translation invariance: accuracies of about 0.8 | Improved |
| add extra augmentation of the same protocol as the test | CNN · CIFAR-10 | translation invariance | Mixed |
| use black fill instead of crop protocol | VGG16, ResNet50, DenseNet121 (PyTorch) · ImageNet 1-pixel shift | P(top-1 change): 0.062 -> 0.18 (VGG16) | Hurt |
| add ReLU (bias=5) after blur and subsampling | toy 1D signal · Table 2 toy example | global sum: 8 and 8 -> 0 and 6 | Hurt |
Closest papers by meaning
- Convolutional Neural Networks Are Not Invariant to Translation, but They Can Learn to Be (JMLR 2021 · similarity 0.90)
- Learning Invariances in Neural Networks from Training Data (NeurIPS 2020 · similarity 0.90)
- Non-Parametric Transformation Networks for Learning General Invariances from Data (AAAI 2019 · similarity 0.90)
- Grounding inductive biases in natural images: invariance stems from variations in data (NeurIPS 2021 · similarity 0.89)
- Do deep networks transfer invariances across classes? (ICLR 2022 · similarity 0.89)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.87)
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.
