Do Convolutional Neural Networks Learn Class Hierarchy?
This addresses the challenge of improving CNN efficiency and accuracy in multi-class image classification, offering incremental insights into learning behavior.
The paper tackles the problem of understanding how Convolutional Neural Networks (CNNs) learn class hierarchies in image classification, finding that confusion patterns follow a hierarchical structure and that early layers separate high-level groups quickly while later layers need more epochs for individual classes, leading to hierarchy-aware CNNs that improve accuracy by accelerating convergence and reducing overfitting.
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.