Research map / Architectures and efficiency / CNN design
Investigating how ReLU-networks encode symmetries
Georg Bökman, Fredrik Kahl
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 horizontal-flip invariance loss | VGG11 · CIFAR10 | GCNN Barrier: 4.00·10^{-2}± 4.9·10^{-3} -> 1.90·10^{-3}± 8.9·10^{-4} | Improved |
| Add invariance loss after 20% of epochs instead of throughout training | VGG11 · CIFAR10 | GCNN Barrier: 1.90·10^{-3}± 8.9·10^{-4} -> 1.33·10^{-2}± 2.8·10^{-3} | Hurt |
| Train without horizontal flipping augmentation | VGG11 · CIFAR10 | GCNN Barrier: 4.00·10^{-2}± 4.9·10^{-3} -> 4.98·10^{-2}± 6.1·10^{-3} | Hurt |
| Add horizontal-flip invariance loss | ResNet50 · ImageNet | GCNN Barrier: 0.0384 -> 0.0054 | Improved |
| Apply horizontal-flip invariance loss | ResNet50 · ImageNet | Accuracy: 0.761 -> 0.754 | Hurt |
| Use image size 225 rather than the default image size | Torchvision old ResNet50 · ImageNet | GCNN Barrier: 0.0384 -> 0.0558 | Hurt |
| Use image size 225 rather than the default image size | Torchvision old ResNet50 · ImageNet | Accuracy: 0.761 -> 0.764 | No clear effect |
| Use image size 225 with invariance loss | Torchvision old ResNet50 · ImageNet | GCNN Barrier: 0.0054 -> 6.72·10^{-5} | Improved |
| Use image size 225 | Simsiam ResNet50 · ImageNet | GCNN Barrier: 0.121 -> 0.103 | Improved |
| Apply REPAIR after merging (ablation) | ResNet50 models · ImageNet | GCNN Barrier: 0.0893 -> 0.0545; 0.0467 -> 0.0384; 0.00636 -> 0.0054; 0.19 -> 0.118; 0.184 -> 0.166; 0.136 -> 0.0805; 0.252 -> 0.121 | Improved |
| Use self-supervised training rather than supervised training | ResNet50 · ImageNet | GCNN Barrier: 0.0384 -> 0.118; 0.0384 -> 0.166; 0.0384 -> 0.0805; 0.0384 -> 0.121 | Hurt |
| Merge a network with its flipped version instead of a separately trained network | ResNet50 · ImageNet | barrier: 20 percentage points; 4.4 percentage points; 2.9 percentage points | Improved |
Closest papers by meaning
- A Characterization Theorem for Equivariant Networks with Point-wise Activations (ICLR 2024 · similarity 0.91)
- A Tale of Two Symmetries: Exploring the Loss Landscape of Equivariant Models (NeurIPS 2025 · similarity 0.90)
- Probing Equivariance and Symmetry Breaking in Convolutional Networks (NeurIPS 2025 · similarity 0.90)
- Lie Group Decompositions for Equivariant Neural Networks (ICLR 2024 · similarity 0.90)
- Learning symmetries via weight-sharing with doubly stochastic tensors (NeurIPS 2024 · similarity 0.90)
- Improving Equivariant Networks with Probabilistic Symmetry Breaking (ICLR 2025 · similarity 0.90)
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.
