High Frequency Component Helps Explain the Generalization of Convolutional Neural Networks
This work addresses the problem of understanding CNN generalization for researchers in machine learning, but it is incremental as it builds on existing observations without introducing new methods or data.
The paper investigates how high-frequency components in images, which are imperceptible to humans, relate to the generalization behavior of convolutional neural networks (CNNs), leading to hypotheses about adversarial examples, robustness-accuracy trade-offs, and training heuristics.
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.