This function is for training a codebook for vector quantization. The data set is split to two clusters, first, and the mean of each cluster is found (centroids). The disttance of each vector from these centroids is found and each vector is associated with a cluster. The mean of vectors of each cluster replaces the centroid first. If the total distance is not improved substantially, The centroids are each split to two. This goes on untill the required number of clusters is reached and the improvement is not substantial.