Learning Vector Quantization: Cluster Size and Cluster Number

We study learning vector quantization methods to adapt the size of (hyper-)spherical clusters to better fit a given data set, especially in the context of non-normalized activations. The basic idea of our approach is to compute a desired radius from the data points that are assigned to a cluster and then to adapt the current radius of the cluster in the direction of this desired radius. Since cluster size adaptation has a considerable impact on the number of clusters needed to cover a data set, we also examine how to select the number of clusters based on validity measures and, in the context of non-normalized activations, on the coverage of the data.