Exploring Generalization in Deep Learning

TL;DR

This paper investigates various explanations for why deep learning models generalize well, analyzing measures like norm control, sharpness, and robustness, and their theoretical and empirical implications.

Contribution

It provides a comprehensive analysis of existing generalization measures, highlighting the role of scale normalization and connecting sharpness with PAC-Bayes theory.

Findings

Scale normalization is crucial for generalization measures.

Sharpness relates to PAC-Bayes bounds.

Different measures explain various phenomena in deep learning.

Abstract

With a goal of understanding what drives generalization in deep networks, we consider several recently suggested explanations, including norm-based control, sharpness and robustness. We study how these measures can ensure generalization, highlighting the importance of scale normalization, and making a connection between sharpness and PAC-Bayes theory. We then investigate how well the measures explain different observed phenomena.