English(EN) On Lipschitz Explosion in Deep Neural Networks with Normalization: Consequences for Optimization and Adversarial Robustness

带归一化的深度神经网络可能出现指数级Lipschitz爆炸

作者 PulseAugur 编辑部 · [1 个来源] · 2026-11-09 05:00

一篇新发表在arXiv上的论文探讨了深度神经网络中“Lipschitz爆炸”的现象，特别是那些包含批量归一化等归一化层的网络。研究表明，即使对每层进行严格的范数控制，这些网络的整体Lipschitz常数也会随着深度的增加呈指数级增长。这种指数级增长可能会破坏优化保证，并使基于Lipschitz的泛化和鲁棒性证书失效。论文还提出，这种爆炸机制会加剧对抗性脆弱性，在MNIST和CIFAR-10等数据集上的实验支持了这样的预测：产生新的奇异方向的扰动更具破坏性。 AI

影响这项研究突显了深度学习模型潜在的脆弱性，特别是在对抗鲁棒性和优化方面，这可能会影响未来的模型设计和安全实践。

排序理由学术论文，详细介绍了深度神经网络的理论发现和实验验证。[lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv stat.ML 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 25 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 学术论文，详细介绍了深度神经网络的理论发现和实验验证。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration: Single-source cluster; Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics: paper, safety; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: Breaking (< 6h); Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

报道来源 [1]

arXiv stat.ML TIER_1 English(EN) · Ashkan Soleymani, Reyhaneh Hosseinpourkhoshkbari, Hadi Daneshmand, Patrick Jaillet · 2026-11-09 05:00

关于归一化深度神经网络中的Lipschitz爆炸：对优化和对抗鲁棒性的影响

arXiv:2611.08775v1 Announce Type: cross Abstract: Can a network look stable at every layer and still be exponentially unstable as a whole? The Lipschitz constant of a neural network is a basic stability quantity appearing in optimization guarantees, generalization bounds, and rob…

本文如何被排名

报道来源 [1]

关于归一化深度神经网络中的Lipschitz爆炸：对优化和对抗鲁棒性的影响

相关实体

相关话题