Dissecting the High-Frequency Bias in Convolutional Neural Networks

For convolutional neural networks (CNNs), a common hypothesis that explains both their generalization capability and their characteristic brittleness is that these models are implicitly regularized to rely on imperceptible high-frequency patterns, more than humans would do. This hypothesis has seen...

Full description

Saved in:
Bibliographic Details
Main Authors: Abello, Antonio A., Hirata, Roberto, Wang, Zhangyang
Format: Conference Proceeding
Language:English
Subjects:
Computer architecture
Computer vision
Conferences
Frequency conversion
Frequency diversity
Pattern recognition
Robustness
Citations: Items that cite this one
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:For convolutional neural networks (CNNs), a common hypothesis that explains both their generalization capability and their characteristic brittleness is that these models are implicitly regularized to rely on imperceptible high-frequency patterns, more than humans would do. This hypothesis has seen some empirical validation, but most works do not rigorously divide the image frequency spectrum. We present a model to divide the spectrum in disjointed discs based on the distribution of energy and apply simple feature importance procedures to test whether high-frequencies are more important than lower ones. We find evidence that mid or high-level frequencies are disproportionately important for CNNs. The evidence is robust across different datasets and networks. Moreover, we find the diverse effects of the network's attributes, such as architecture and depth, on frequency bias and robustness in general. Code for reproducing our experiments is available at: https://github.com/Abello966/FrequencyBiasExperiments
ISSN:2160-7516
DOI:10.1109/CVPRW53098.2021.00096