On the Role of Hierarchy for Neural Network Interpretation

Abstract

In this paper, we concentrate on the expressive power of hierarchical structures in neural networks. Recently, the so-called SplitNet model was introduced. It develops a dynamic network structure based on growing and splitting Kohonen chains and it belongs to the class of topology preserving networks. We briefly introduce the basics of this model and explain the different sources of information built up during the training phase, namely the neuron distribution, the final topology of the network, and the emerging hierarchical structure. In contrast to most other neural models in which the structure is only a means to get desired results, in SplitNet the structure itself is part of the aim . Our focus then lies on the interpretation of the hierarchy produced by the training algorithm and we relate our findings to a common data analysis method, the hierarchical cluster analysis. We illustrate the results of network application to a real medical diagnosis and monitoring task in the domain of nerve lesions of the human hand.