Research map / Architectures and efficiency / CNN design
Preventing Gradient Attenuation in Lipschitz Constrained Convolutional 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 BCOP orthogonal convolutions rather than OSSN, RKO, or SVCM | Lipschitz-constrained Small and Large image classifiers · MNIST and CIFAR10 | robust accuracy: 42.95\pm1.09 / 43.58\pm0.44 / 28.94\pm1.58 -> 45.84\pm0.90; 53.77\pm1.02 / 55.18\pm0.46 / 38.00\pm1.82 -> 56.37\pm0.33; | Helped |
| Use BCOP instead of competing Lipschitz convolutions | Wasserstein critic with MaxMin activation · STL-10 GAN | Wasserstein distance estimate: 7.39\pm0.31 / 8.95\pm0.12 -> 9.91\pm0.11 | Helped |
| Use BCOP instead of competing Lipschitz convolutions | Wasserstein critic with ReLU activation · STL-10 GAN | Wasserstein distance estimate: 7.06\pm0.72 / 7.82\pm0.21 -> 8.28\pm0.19 | Helped |
| Train convolution weights rather than freezing them (ablation) | BCOP image classifier · MNIST and CIFAR10, Small and Large | robust accuracy: 7.51\pm1.18 -> 45.84\pm0.90; 0.00\pm0.00 -> 56.37\pm0.33; 36.44\pm0.70 -> 50.01\pm0.21; 27.47\pm1.37 -> 58.26\pm0.17 | Helped |
| Replace BCOP with RK-L2NE convolution bounds (ablation) | Lipschitz-constrained image classifier · MNIST and CIFAR10, Small and Large | robust accuracy: 45.84\pm0.90 vs 39.77\pm0.73; 56.37\pm0.33 vs 37.79\pm1.21; 50.01\pm0.21 vs 44.65\pm0.61; 58.26\pm0.17 vs 43.40\pm0.46 | Hurt |
Closest papers by meaning
- Improved techniques for deterministic l2 robustness (NeurIPS 2022 · similarity 0.93)
- Enhancing Certified Robustness via Block Reflector Orthogonal Layers and Logit Annealing Loss (ICML 2025 · similarity 0.89)
- 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)
- DP-SGD Without Clipping: The Lipschitz Neural Network Way (ICLR 2024 · similarity 0.88)
- Approximation theory for 1-Lipschitz ResNets (NeurIPS 2025 · similarity 0.88)
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.