Self-Organizing Maps and Unsupervised Classification - Neural Networks: Methodology and Applications - page 400

Fig. 7.10. Performing batch algorithm with T fixed ( from top to bottom and from
left to right : T = 10, T =5, T =3and T =1)
Figure 7.11 shows that the cooling schedule may be crucial. On that figure,
several final results are displayed for the same training data set, the same
temperature variation interval and different cooling schedules. If temperature
decreases too fast (quenching), the self-organization is not e cient and the
induced neighborhood relation among reference vectors from the data space
is not a faithful representation of the graph-induced topology. Using neuronal
analogy, one may say that the receptive fields of neurons that are close in
cortical maps are not close in the perceptual space. Note that topological self-
organization is highly sensitive to the whole parameter set of the algorithm. No