Research map / Architectures and efficiency / CNN design
Convolutional Neural Networks Are Not Invariant to Translation, but They Can Learn to Be
Valerio Biscione, Jeffrey S. Bowers
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 |
|---|---|---|---|
| pretrain on ImageNet before fine-tuning on one location | CNNs (ResNet-50, VGG16, VGG16bn, AlexNet, DenseNet-121) · one-location datasets, fully-translated test | accuracy | Improved |
| pretrain on ImageNet | VGG16 · EMNIST | accuracy (%): 5±0.76 -> 84±3.12 | Improved |
| add global average pooling to untrained VGG16/VGG16bn (ablation) | VGG16, VGG16bn · one-location datasets | accuracy (%): ~98 | Improved |
| global average pooling already present in ResNet-50 (ablation) | untrained ResNet-50 · one-location datasets | accuracy | No clear effect |
| add batch normalization (ablation) | untrained VGG16 · one-location datasets | accuracy | Improved |
| add batch normalization (ablation) | ImageNet-pretrained VGG16 · one-location datasets (all but LEEK10) | accuracy | Hurt |
| add anti-aliased max pooling or fully-convolutional padding (ablation) | VGG16bn · KMNIST, MNIST, other one-location datasets | accuracy | Mixed |
| pretrain on fully-translated dataset instead of one-location fine-tune only | ResNet-50, VGG16, VGG16bn, AlexNet · Invariance Metric on novel datasets | Invariance Metric: < 0.2 -> 0.83±0.1 | Improved |
| fine-tune on one-location data after translated pretraining on same dataset | CNN pretrained on fully-translated ShapeNet · ShapeNet fully-translated test | accuracy (%): at least 95 -> 79.56 | Hurt |
| pretrain on a more complex translated dataset than the fine-tuning one (ablation) | CNNs · cross-dataset fine-tuning | accuracy | Improved |
| train on whole canvas with classes segregated by area (no full translation) (ablation) | VGG16 · EMNIST18 / MNIST | accuracy | No clear effect |
| leave upper-right quadrant empty during pretraining (ablation) | VGG16 · EMNIST | accuracy on untrained area | Mixed |
Closest papers by meaning
- Lost in Translation: Modern Neural Networks Still Struggle With Small Realistic Image Transformations (ECCV 2024 · similarity 0.90)
- Why do deep convolutional networks generalize so poorly to small image transformations? (JMLR 2019 · similarity 0.90)
- On Translation Invariance in CNNs: Convolutional Layers Can Exploit Absolute Spatial Location (CVPR 2020 · similarity 0.88)
- Making Sense of CNNs: Interpreting Deep Representations & Their Invariances with INNs (ECCV 2020 · similarity 0.88)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.88)
- B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable (NeurIPS 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.
