High Frequency Component Helps Explain the Generalization of Convolutional Neural Networks

TL;DR

This paper explores how CNNs' ability to capture high-frequency image components, which are imperceptible to humans, influences their generalization, robustness, and training behaviors.

Contribution

It provides insights into the role of high-frequency components in CNN generalization and offers hypotheses linking frequency spectrum to adversarial vulnerability and training heuristics.

Findings

CNN captures high-frequency image components.

High-frequency components relate to adversarial examples.

Trade-off between robustness and accuracy in CNNs.

Abstract

We investigate the relationship between the frequency spectrum of image data and the generalization behavior of convolutional neural networks (CNN). We first notice CNN's ability in capturing the high-frequency components of images. These high-frequency components are almost imperceptible to a human. Thus the observation leads to multiple hypotheses that are related to the generalization behaviors of CNN, including a potential explanation for adversarial examples, a discussion of CNN's trade-off between robustness and accuracy, and some evidence in understanding training heuristics.