The weights of a few of the nodes surrounding the winning node are also adjusted using eqn. 7. These nodes are described as being within the winning node's 'neighbourhood', and the degree to which their weight vectors are adjusted is governed by the neighbourhood function, h_j,k(t). This is determined using the following equation (Kohonen 1998):

where d_j,k is the euclidean distance between nodes j and k
  s(t) is the neighbourhood width given by the equation (Haykin, 1999):

here, s₀ is the initial neighbourhood width. The time constant t₂ controls the rate at which s falls.
It is suggested that the initial value of h be 0.1, and a value of t₁ be chosen in order that h always stays above 0.01. s₀ should be in the order of magnitude of the width of the grid, and t₂ be chosen in order that h_j,i for adjacent nodes eventually decreases to zero.

3.5.1. Learning Vector Quantisation

After the clustering phase has taken place, the map can be further fine tuned by re-introducing a small number of the training patterns where the type is known. Each pattern is again presented to the SOM, and the weight of the winning node adjusted depending on whether the pattern was correctly identified or not.
For correctly identified patterns (Kohonen 1998):

For incorrectly identified patterns:

  where w_n is the weight vector of the winning node n at time t.