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. Our
observation also immediately inspire methods related to the adversarial attack
and defense methods.
Description
[1905.13545] High Frequency Component Helps Explain the Generalization of Convolutional Neural Networks
%0 Generic
%1 wang2019frequency
%A Wang, Haohan
%A Wu, Xindi
%A Huang, Zeyi
%A Xing, Eric P.
%D 2019
%K 2019 cnn deep-learning
%T High Frequency Component Helps Explain the Generalization of
Convolutional Neural Networks
%U http://arxiv.org/abs/1905.13545
%X 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. Our
observation also immediately inspire methods related to the adversarial attack
and defense methods.
@misc{wang2019frequency,
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. Our
observation also immediately inspire methods related to the adversarial attack
and defense methods.},
added-at = {2020-01-07T19:58:19.000+0100},
author = {Wang, Haohan and Wu, Xindi and Huang, Zeyi and Xing, Eric P.},
biburl = {https://www.bibsonomy.org/bibtex/2f60917704cf78781b612c3430d2c3121/analyst},
description = {[1905.13545] High Frequency Component Helps Explain the Generalization of Convolutional Neural Networks},
interhash = {3756e3a9e44005276d47f3faec5a2a4d},
intrahash = {f60917704cf78781b612c3430d2c3121},
keywords = {2019 cnn deep-learning},
note = {cite arxiv:1905.13545},
timestamp = {2020-01-07T19:58:19.000+0100},
title = {High Frequency Component Helps Explain the Generalization of
Convolutional Neural Networks},
url = {http://arxiv.org/abs/1905.13545},
year = 2019
}