LVQ2.

The version described here as LVQ2 corresponds actually the LVQ2.1 [Kohonen, 1990a]. For each stochastic input sample $\mbox{\boldmath$x$} \in {\cal R}^D$,the adjustments are performed for the two best-matching codebook vectors $\mbox{\boldmath$m$}_c$ and $\mbox{\boldmath$m$}_{c'}$,which are found using the minimum Euclidean distance criterion (1).

If $\mbox{\boldmath$m$}_c$ and $\mbox{\boldmath$x$}$ belong to the same class, but $\mbox{\boldmath$m$}_{c'}$ and $\mbox{\boldmath$x$}$ to different classes, respectively, then

where the teaching gain $\alpha(t) \in ]0,1[$ is decreased monotonically. An additional constraint for the update is that the sample $\mbox{\boldmath$x$}$ should lie in a narrow window between $\mbox{\boldmath$m$}_c$ and $\mbox{\boldmath$m$}_{c'}$ [Kohonen, 1995].