Study finds normalization layers can mask exponential instability in deep networks

A new arXiv paper shows deep neural networks with normalization can appear stable layer-by-layer while their overall Lipschitz constant grows exponentially, affecting optimization and adversarial robustness.

Read the full story at arXiv cs.LG ↗

Timeline · 1 report

1. 2026-11-09 05:00 · arXiv cs.LG
   On Lipschitz Explosion in Deep Neural Networks with Normalization: Consequences for Optimization and Adversarial Robustness