Unsupervised ontogenic networks
Bernd Fritzke · 2004
In this section, network models are described which learn unsupervised and generate their topology during learning. Among the models described, one can distinguish those having a specific dimensionality (e.g. two or three) from models whose dimensionality varies locally with the input data. Furthermore, models with a fixed number of units but variable connectivity can be distinguished from models which also change their number of units through insertion and/or deletion. Application areas of the described models include vector quantization, data visualization and clustering. C2.4.1 Introduction First, an incremental self-organizing network model from the author is described which is called `growing cell structures'. The model is able to generate dimensionality reducing mappings which may be used, for example, for visualization of high-dimensional data or for clustering. The basic building blocks of the generated topology are hypertetrahedrons of a certain dimensionality k chosen...