AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Investigating how ReLU-networks encode symmetries

Georg Bökman, Fredrik Kahl

NeurIPS 2023 · Architectures and efficiency › CNN design · approach: Group symmetry equivariant networks · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
Add horizontal-flip invariance lossVGG11 · CIFAR10GCNN 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 trainingVGG11 · CIFAR10GCNN Barrier: 1.90·10^{-3}± 8.9·10^{-4} -> 1.33·10^{-2}± 2.8·10^{-3}Hurt
Train without horizontal flipping augmentationVGG11 · CIFAR10GCNN Barrier: 4.00·10^{-2}± 4.9·10^{-3} -> 4.98·10^{-2}± 6.1·10^{-3}Hurt
Add horizontal-flip invariance lossResNet50 · ImageNetGCNN Barrier: 0.0384 -> 0.0054Improved
Apply horizontal-flip invariance lossResNet50 · ImageNetAccuracy: 0.761 -> 0.754Hurt
Use image size 225 rather than the default image sizeTorchvision old ResNet50 · ImageNetGCNN Barrier: 0.0384 -> 0.0558Hurt
Use image size 225 rather than the default image sizeTorchvision old ResNet50 · ImageNetAccuracy: 0.761 -> 0.764No clear effect
Use image size 225 with invariance lossTorchvision old ResNet50 · ImageNetGCNN Barrier: 0.0054 -> 6.72·10^{-5}Improved
Use image size 225Simsiam ResNet50 · ImageNetGCNN Barrier: 0.121 -> 0.103Improved
Apply REPAIR after merging (ablation)ResNet50 models · ImageNetGCNN 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 -> 0.121Improved
Use self-supervised training rather than supervised trainingResNet50 · ImageNetGCNN Barrier: 0.0384 -> 0.118; 0.0384 -> 0.166; 0.0384 -> 0.0805; 0.0384 -> 0.121Hurt
Merge a network with its flipped version instead of a separately trained networkResNet50 · ImageNetbarrier: 20 percentage points; 4.4 percentage points; 2.9 percentage pointsImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy