To ease understanding, the above algorithm represents
the index of codewords in one-dimension (0…h - 1).
Actually, the index can simply be extended to twodimensions
which is usually used in practice.
At this point the most appropriate form of the neighbourhood
update function can be decided. According to
Table 3, it has been preliminarily concluded that the linearly decaying neighbourhood update function with
k = 2 is as good as the squared decay and the complex
Gaussian function with k= -0.125. This form of update
function is very appropriate for use in conjunction with
Eqs. 12–14 since it is equivalent to increasing the number
of right shift time of the denominator term in both equations.
For a neighbour whose index is j far away from the
winner neuron, only the denominator term of Eq. 11 needs
to be updated. This changes the form of the denominator
term in Eq. 13 to be: