Researchers · Interviewed at NeurIPS 2025

Linara Adilova

TU Dortmund

NeurIPS 2025 in San Diego. Deep Learning Secrets

NeurIPS 2025 · December 23, 2025 · 8 min

NeurIPS 2025 in San Diego. Deep Learning Secrets

In this post #NeurIPS2025 episode, Linara Adilova shares the work on relative flatness in deep learning — a concept that could explain why neural networks generalize so well.

Key insights: • How relative flatness differs from neural collapse • Experiments showing why flatness may be necessary for generalization • The layer-wise behavior of networks and what it means for future AI research

More from NeurIPS 2025

All event interviews →