Research map / Learning theory and algorithms / Kernels and regression theory
Spectrally-normalized margin bounds for neural 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 |
|---|---|---|---|
| train on randomized labels instead of true labels | AlexNet · CIFAR10 | normalized margin distribution | Hurt |
| train on randomized labels instead of true labels | AlexNet · MNIST | normalized margin distribution | Hurt |
| use CIFAR100 labels instead of CIFAR10 labels for the same images | AlexNet · CIFAR100 vs. CIFAR10 with randomized labels | normalized margin distribution | No clear effect |
| replace CIFAR10 images with random images sampled from matching Gaussian image statistics | AlexNet · CIFAR10 | normalized margin distribution | Hurt |
| add l2 weight decay (ablation) | AlexNet · CIFAR10 | normalized margin distribution | No clear effect |
| add l2 weight decay (ablation) | AlexNet · CIFAR10 | generalization | No clear effect |
| use l1, (2,1), or l2 norms in the spectral-complexity correction term (ablation) | spectrally-normalized margin analysis · MNIST, CIFAR10, and CIFAR100 experiments | margin distributions | No clear effect |
| use low-margin rather than high-margin CIFAR10 inputs | AlexNet predictions under uniform input noise at scale 0.15 · CIFAR10 | label-flip rate: 0.04% -> 5.86% | Hurt |
| normalize margins by spectral complexity | AlexNet trained across epochs · CIFAR10 | margin distribution over epochs | Helped |
| train on MNIST rather than CIFAR10 | AlexNet · MNIST vs. CIFAR10 | normalized margin distribution | Helped |
Closest papers by meaning
- Data-dependent Sample Complexity of Deep Neural Networks via Lipschitz Augmentation (NeurIPS 2019 · similarity 0.88)
- Some Fundamental Aspects about Lipschitz Continuity of Neural Networks (ICLR 2024 · similarity 0.86)
- Fantastic Generalization Measures are Nowhere to be Found (ICLR 2024 · similarity 0.86)
- ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks (NeurIPS 2024 · similarity 0.86)
- Generalization Bounds for Rank-sparse Neural Networks (NeurIPS 2025 · similarity 0.86)
- Generalizability of Neural Networks Minimizing Empirical Risk Based on Expressive Power (ICLR 2025 · similarity 0.85)
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.