Do Convolutional Neural Networks Learn Class Hierarchy?

TL;DR

This paper investigates how CNNs inherently learn class hierarchies, revealing that early layers capture high-level groupings quickly, which can be exploited to improve training efficiency and data quality assessment.

Contribution

The study introduces visual-analytics methods to uncover class hierarchies in CNNs and demonstrates how leveraging this hierarchy enhances model convergence and data quality diagnostics.

Findings

Early CNN layers detect high-level class groups rapidly.

Hierarchy-aware CNNs improve training speed and reduce overfitting.

Methods identify training data quality issues.

Abstract

Convolutional Neural Networks (CNNs) currently achieve state-of-the-art accuracy in image classification. With a growing number of classes, the accuracy usually drops as the possibilities of confusion increase. Interestingly, the class confusion patterns follow a hierarchical structure over the classes. We present visual-analytics methods to reveal and analyze this hierarchy of similar classes in relation with CNN-internal data. We found that this hierarchy not only dictates the confusion patterns between the classes, it furthermore dictates the learning behavior of CNNs. In particular, the early layers in these networks develop feature detectors that can separate high-level groups of classes quite well, even after a few training epochs. In contrast, the latter layers require substantially more epochs to develop specialized feature detectors that can separate individual classes. We demonstrate how these insights are key to significant improvement in accuracy by designing hierarchy-aware CNNs that accelerate model convergence and alleviate overfitting. We further demonstrate how our methods help in identifying various quality issues in the training data.