Showing papers by "Christian Blum published in 1997"

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 spl i t t ing 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 bui l t up during the training phase, namely the neuron distr ibut ion, 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 il lustrate the results of network application to a real medical diagnosis and monitoring task in the domain of nerve lesions of the human hand.

Interpretation of a Hierarchical Neural Network

Abstract: In this paper, we concentrate on the expressive power of hierarchical structures in data analysis. Recently, the so-called Split Net model was introduced. It develops a dynamic, growing network structure and 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. Our focus then lies on the interpretation of the hierarchy produced by the training algorithm and we relate our findings to existing data analysis methods. We illustrate the results with an example from a real medical diagnosis and monitoring task.

Diagnosis and Monitoring of Ulnar Nerve Lesions

Abstract: In this paper we introduce a novel approach for diagnosis and monitoring of ulnar nerve lesions, affecting the coordination of movement of the ring and little finger of the human hand. Based on data generated by ultrasound measurements, we developed suitable preprocessing methods for automatic extraction of relevant features from the movement pattern to be examined. The partial absence of class information even for the pattern in the training set requires the use of unsupervised methods for the learning and class assignment procedures. For that reason, we use a new dynamic and hierarchic neural network for the analysis of the generated pattern vectors. The dynamically structured architecture of the network satisfies the special needs of this medical task, such as providing variable levels of generalization and efficient retrieval of similar cases.