Detailed guidelines for how to actually compute the maps are given by Kohonen (1995c), as well as in the documentation of the public domain program package SOM_PAK [Kohonen et al., 1996a]. The reference vectors are first initialized to lie in an ordered configuration on the plane spanned by the two principal eigenvectors of the data, and thereafter taught in a two-phase process. The learning starts with a wide neighborhood kernel covering most of the map, and during the first phase the kernel quickly narrows close to its final width, at the same time becoming smaller in its peak amplitude. During the second, longer phase, the neighborhood kernel continues from the narrower form and slowly shrinks to its final width and magnitude. The first phase enforces a global ordering of the map, while in the second phase the final accurate state of the map is formed gradually. The final neighborhood width determines the ``stiffness'' of the map, i.e., how closely the map will follow the local structures in the data.