Neural Maps and Learning Vector Quantization for Faithful Data Mining in Medical Data Analysis
Dr. Thomas Villmann
Head of the computational intelligence group
University Leipzig / Germany, Hospital for Psychotherapy

Abstract:

Neural Maps are special artificial neural networks which are adapted from the cortex in real brains. The cortex processes the sensoric information at a first level. Thereby, the information flow is optimized by data driven adaptation of the several cortex areas responsible for different stimuli. Neural maps transfer these functional views into a technical context of artificial neural networks for data mining and representation. We will consider several properties and variants of neural maps for faithful data analysis. In particular we will concentrate on the self-organizing map model (SOM), which generates under certain conditions a topology preserving map, i.e. a low-dimensional representation of high-dimensional data can be achieved. We discuss useful extensions of the basic SOM, such as growing variants and information optimum coding for faithful data modeling. We provide tools to assess the quality of topology preservation of the map, which is necessary for correct interpretation.