Research map / Learning theory and algorithms / Combinatorial and search optimization
Efficiently Computing Local Lipschitz Constants of Neural Networks via Bound Propagation
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 tight linear relaxation for Clarke gradients instead of RecurJac's loose interval relaxation (ablation) | bound propagation Lipschitz method · MNIST 3-layer MLP | local Lipschitz bound: 1,091.31 -> 688.15 | Helped |
| same, on CNN (ablation) | bound propagation Lipschitz method · MNIST CNN-2C2F | local Lipschitz bound: 12,514.55 -> 5,473.03 | Helped |
| same, on CIFAR-10 CNN-4C2F (ablation) | bound propagation Lipschitz method · CIFAR-10 CNN-4C2F | local Lipschitz bound: 12502332.00 -> 1049447.88 | Helped |
| same, on TinyImageNet CNN-4C2F (ablation) | bound propagation Lipschitz method · TinyImageNet CNN-4C2F | local Lipschitz bound: 714189.12 -> 25096.26 | Helped |
| add branch-and-bound (ablation) | Ours (w/o BaB) · MNIST 3-layer MLP | local Lipschitz bound: 688.15 -> 397.25 | Helped |
| add branch-and-bound (ablation) | Ours (w/o BaB) · MNIST CNN-2C2F | local Lipschitz bound: 5,473.03 -> 5,458.84 | No clear effect |
| add branch-and-bound (ablation) | Ours (w/o BaB) · Synthetic MLP width 128 | local Lipschitz bound: 35.66 -> 28.69 | Helped |
| add branch-and-bound (ablation) | Ours (w/o BaB) · Synthetic MLP width 32 | local Lipschitz bound: 12.28 -> 12.13 | No clear effect |
| replace RecurJac relaxation with ours for monotonicity verification (ablation) | Adult MLP · Adult monotonicity (age) | % examples verified: 32% -> 40% | Helped |
Closest papers by meaning
- Exactly Computing the Local Lipschitz Constant of ReLU Networks (NeurIPS 2020 · similarity 0.94)
- ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks (NeurIPS 2024 · similarity 0.92)
- Some Fundamental Aspects about Lipschitz Continuity of Neural Networks (ICLR 2024 · similarity 0.90)
- Novel Quadratic Constraints for Extending LipSDP beyond Slope-Restricted Activations (ICLR 2024 · similarity 0.88)
- Approximation theory for 1-Lipschitz ResNets (NeurIPS 2025 · similarity 0.88)
- A provable control of sensitivity of neural networks through a direct parameterization of the overall bi-Lipschitzness (NeurIPS 2024 · 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.