ConfusionTree-Pattern: A Hierarchical Design for an Efficient and Performant Multi-Class Pattern

Abstract

Developing neural networks for supervised multi-class classification has become important for theory and practice. An essential point is the design of the underlying network. Beside single-network approaches there are several multi-class patterns which decompose a classification problem into multiple sub-problems and derive systems of neural networks. We show that existing multi-class patterns can be improved by a new and simple labeling scheme for the training of the sub-problems. We efficiently derive a class hierarchy which is optimized for our labeling scheme and, unlike most of existing works, has no schematic restrictions. Based on that we introduce a hierarchical multi-class pattern, called ConfusionTree-pattern, which is able to reach high classification accuracies. Our experiments show that our multi-class ConfusionTree-pattern reaches state-of-the-art results regarding performance and efficiency.