Generalization in Deep Learning

TL;DR

This paper offers theoretical explanations for the generalization ability of deep learning models, addressing open questions and discussing approaches to establish non-vacuous guarantees, while highlighting limitations and future research directions.

Contribution

It provides new theoretical insights into deep learning generalization, proposes methods for non-vacuous guarantees, and identifies open problems and limitations.

Findings

Deep learning can generalize well despite high capacity.

Theoretical approaches for non-vacuous generalization guarantees.

Discussion of limitations and future research directions.

Abstract

This paper provides theoretical insights into why and how deep learning can generalize well, despite its large capacity, complexity, possible algorithmic instability, nonrobustness, and sharp minima, responding to an open question in the literature. We also discuss approaches to provide non-vacuous generalization guarantees for deep learning. Based on theoretical observations, we propose new open problems and discuss the limitations of our results.