Abstract:
This paper presents two variations of the Self-Organizing Map in order to adapt special input signal attributes. The first one improves the properties concerning outliers and clusters with low feature density within the provided training data set by changing the neighbourhood function while training the net. In the second introduced modification the originally two- dimensional feature map has been extended to an ordered two- dimensional structure of two-dimensional feature maps. The Quasi-Four-Dimensional Neuroncube (QFDN) is introduced. To each variation a short description of a prospective application has been attached.