Do Convolutional Neural Networks Learn Class Hierarchy?

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) achieve high accuracy in image classification.
• Accuracy declines with increasing class numbers due to confusion.
• Class confusion in CNNs exhibits an underlying hierarchical structure.

Purpose of the Study:

• To develop visual-analytics methods for revealing and analyzing the hierarchy of similar classes.
• To investigate the relationship between class hierarchy and CNN internal data.
• To leverage hierarchical insights for improving CNN performance and data quality.

Main Methods:

• Utilized visual-analytics techniques to explore CNN internal representations.
• Analyzed the hierarchical structure of class confusions.
• Correlated class hierarchy with the learning behavior of CNNs across different layers.

Main Results:

• The class hierarchy dictates both confusion patterns and CNN learning dynamics.
• Early CNN layers learn high-level class group features rapidly.
• Later CNN layers require more epochs for specialized feature detection.

Conclusions:

• Hierarchy-aware CNN designs accelerate convergence and reduce overfitting.
• The developed methods enhance image classification accuracy.
• Visualizing class hierarchy aids in identifying training data quality issues.