Research map / Architectures and efficiency / CNN design
Direct Parameterization of Lipschitz-Bounded Deep Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use Sandwich layers instead of AOL, Orthogonal, or SLL parameterizations | Lipschitz-bounded network · Toy square-wave fitting, γ=10 | Lip. tightness (%): 67.9 -> 94.0 | Helped |
| Use Sandwich layers instead of SLL | Lipschitz-bounded network · Toy square-wave fitting, γ=5 | Lip. tightness (%): 90.5 -> 99.3 | Helped |
| Use Sandwich layers instead of SLL Large | CIFAR-100 classifier · CIFAR-100 empirical robustness, ε=72/255 | AutoAttack accuracy: 34.9 -> 40.2 | Helped |
| Use Sandwich Large instead of SLL Large | CIFAR-100 classifier · CIFAR-100 clean | accuracy: 54.8 -> 57.5 | Helped |
| Use Sandwich Medium instead of SLL Medium | TinyImageNet classifier · TinyImageNet empirical robustness, ε=72/255 | AutoAttack accuracy: 19.8 -> 25.3 | Helped |
| Use Sandwich Medium instead of SLL Medium | TinyImageNet classifier · TinyImageNet clean | accuracy: 30.3 -> 35.0 | Helped |
| Use Sandwich instead of SLL XLarge | CIFAR-100 classifier · CIFAR-100 certified robustness, ε=36/255 | certified accuracy: 36.5 -> 35.3 | Hurt |
| Use Sandwich instead of SLL XLarge | TinyImageNet classifier · TinyImageNet certified robustness, ε=36/255 | certified accuracy: 23.0 -> 24.7 | Helped |
| Increase the prescribed Lipschitz bound from γ=1 to γ=10 | Sandwich curve-fitting model · Toy square-wave fitting | Lip. tightness (%): 99.9 -> 94.0 | Hurt |
| Use a smaller Lipschitz bound | Lipschitz-bounded image classifier · CIFAR-10/100, large attack sizes ε=72/255 or 108/255 | robust test accuracy | Helped |
| Choose a Lipschitz bound that is too small | Lipschitz-bounded image classifier · CIFAR-10/100 | clean test accuracy | Hurt |
| Apply Lipschitz-bounded models rather than an unregularized CNN | Image classifier · CIFAR-10/100 training curves | test error over epochs | Helped |
Closest papers by meaning
- Preventing Gradient Attenuation in Lipschitz Constrained Convolutional Networks (NeurIPS 2019 · similarity 0.90)
- DP-SGD Without Clipping: The Lipschitz Neural Network Way (ICLR 2024 · similarity 0.90)
- ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks (NeurIPS 2024 · similarity 0.89)
- A Rescaling-Invariant Lipschitz Bound Based on Path-Metrics for Modern ReLU Network Parameterizations (ICML 2025 · similarity 0.89)
- Novel Quadratic Constraints for Extending LipSDP beyond Slope-Restricted Activations (ICLR 2024 · similarity 0.87)
- Some Fundamental Aspects about Lipschitz Continuity of Neural Networks (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.