Research map / Architectures and efficiency / RNNs and normalization
Learning Layer-wise Equivariances Automatically using Gradients
Tycho F.A. van der Ouderaa, Alexander Immer, Mark van der Wilk
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 |
|---|---|---|---|
| train with Laplace marginal likelihood instead of MAP | F-FC+CONV · CIFAR-10 test | test accuracy: 81.61 -> 83.92 | Improved |
| train with Laplace marginal likelihood instead of MAP | FC+CONV · CIFAR-10 test | test accuracy: 76.93 -> 83.32 | Improved |
| train with Laplace marginal likelihood instead of MAP | S-FC+S-CONV (prior on u) · CIFAR-10 test | test accuracy: 58.80 -> 81.55 | Improved |
| train with Laplace marginal likelihood instead of MAP | CONV · CIFAR-10 test | test accuracy: 82.81 -> 84.07 | Improved |
| train with Laplace marginal likelihood instead of MAP | S-CONV (prior on omega) · CIFAR-10 test | test accuracy: 81.84 -> 81.20 | Hurt |
| add learnable symmetry (FC path with CONV) versus fixed CONV | F-FC+CONV vs CONV · CIFAR-10 test | test accuracy: 84.07 -> 83.92 | No clear effect |
| use learnable symmetry model instead of strict CONV | F-FC+CONV · Quadrant MNIST digit+quadrant | test accuracy: 24.50 -> 99.00 | Improved |
| use learnable symmetry model instead of FC | F-FC+CONV · Quadrant MNIST digit+quadrant | test accuracy: 95.46 -> 99.00 | Improved |
| use learnable symmetry instead of strict CONV on strictly symmetric task | F-FC+CONV · Quadrant MNIST digit | test accuracy: 99.23 -> 98.86 | No clear effect |
| Laplace marginal likelihood for fully connected net instead of MAP | FC · CIFAR-10 test | test accuracy: 64.42 -> 52.36 | Hurt |
| add rotation-equivariant GCONV path, train with Laplace | F-FC+CONV+GCONV · CIFAR-10 | test accuracy: 68.33 -> 80.94 | Improved |
| add GCONV path, train with Laplace | F-FC+CONV+GCONV · Rotated CIFAR-10 | test accuracy: 48.30 -> 55.68 | Improved |
Closest papers by meaning
- A Tale of Two Symmetries: Exploring the Loss Landscape of Equivariant Models (NeurIPS 2025 · similarity 0.91)
- Probing Equivariance and Symmetry Breaking in Convolutional Networks (NeurIPS 2025 · similarity 0.91)
- Learning (Approximately) Equivariant Networks via Constrained Optimization (NeurIPS 2025 · similarity 0.91)
- On the hardness of learning under symmetries (ICLR 2024 · similarity 0.91)
- A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs (ICML 2024 · similarity 0.90)
- Bridging Symmetry and Robustness: On the Role of Equivariance in Enhancing Adversarial Robustness (NeurIPS 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.
