Research map / Architectures and efficiency / CNN design
Almost-Orthogonal Layers for Efficient General-Purpose Lipschitz Networks
Bernd Prach, Christoph H. Lampert
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 |
|---|---|---|---|
| increase loss margin offset u and temperature t from sqrt2/16 to 16 sqrt2 (u scaled with t) (ablation) | AOL-Small · CIFAR-10 | standard accuracy: 79.8% -> 48.2% | Hurt |
| increase loss margin offset u and temperature t from sqrt2/16 to 4 sqrt2 (ablation) | AOL-Small · CIFAR-10 (eps=1) | certified robust accuracy: 0.0% -> 30.8% | Improved |
| increase margin from u=sqrt2/4 to u=sqrt2 (default) (ablation) | AOL-Small · CIFAR-10 (eps=108/255) | certified robust accuracy: 33.0% -> 47.9% | Improved |
| increase margin from u=4 sqrt2 to u=16 sqrt2 (ablation) | AOL-Small · CIFAR-10 (eps=1) | certified robust accuracy: 30.8% -> 28.6% | Hurt |
| increase network size from AOL-Small to AOL-Large | AOL · CIFAR-10 (eps=36/255) | certified robust accuracy: 62.0% -> 64.0% | Improved |
| increase network size from AOL-Small to AOL-Large | AOL · CIFAR-10 (eps=1) | certified robust accuracy: 21.8% -> 23.7% | Improved |
| increase network size from AOL-Small to AOL-Large | AOL · CIFAR-100 (eps=1) | certified robust accuracy: 6.7% -> 7.8% | Improved |
| use AOL-ALT architecture instead of AOL-STD (ablation) | AOL · CIFAR-10 (eps=36/255) | certified robust accuracy: 56.4% -> 60.3% | Improved |
| use AOL-DIL (dilated, circular padding) architecture instead of AOL-ALT (ablation) | AOL · CIFAR-10 (eps=36/255) | certified robust accuracy: 60.3% -> 54.2% | Hurt |
| use fully connected AOL-FC architecture instead of AOL-STD (ablation) | AOL · CIFAR-10 (eps=1) | certified robust accuracy: 15.8% -> 17.6% | Improved |
| AOL-Large compared with SOC-15 + CR at large threshold | AOL-Large · CIFAR-10 (eps=108/255) | certified robust accuracy: 35.5% -> 49.0% | Improved |
| AOL-Large compared with SOC-15 + CR on clean accuracy | AOL-Large · CIFAR-10 | standard accuracy: 76.4% -> 71.6% | Hurt |
Closest papers by meaning
- A Unified Algebraic Perspective on Lipschitz Neural Networks (ICLR 2023 · similarity 0.93)
- Enhancing Certified Robustness via Block Reflector Orthogonal Layers and Logit Annealing Loss (ICML 2025 · similarity 0.90)
- Approximation theory for 1-Lipschitz ResNets (NeurIPS 2025 · similarity 0.89)
- ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks (NeurIPS 2024 · similarity 0.88)
- A Rescaling-Invariant Lipschitz Bound Based on Path-Metrics for Modern ReLU Network Parameterizations (ICML 2025 · similarity 0.87)
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · 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.
