sturbing hidden units.
The above addressed problem may not notably deteriorate
the performance of CL on an application of vector quantization, where the goal is not to find any clusters or classes.
However,even in this case the input data density is usually not
uniform, and results like the one shown in Fig. l(a) are desired.
For the codebook we would now have three weight vectors,
each located at the center of one of the clusters. Results like
that shown in Fig. l(c) will increase the number of vectors
in the codebook, without much improvement on the coding
performance due to the fact that the code vectors which are
located at the boundary points between clusters can contribute
only a little in reducing the distortion.
Therefore, we see that the conscience strategy and FSCL
work well only when the number of clusters in the input data
set is known in advance so that we can let our CL net have
the same number of units. This is not an easy task since we
usually do not know the number of clusters in the input data a
priori. The same problem exists in the conventional k-means
clustering method: if the number of clusters k is selected
inappropriately, we may obtain very poor clustering results.
Unfortunately, the selection of k is a hard problem. It could
only be solved heuristically by some prior knowledge, or by
enumerating a number of different values and doing clustering
for each of these values so that a better value could be obtained
according to some rule, e.g., finding the value with a sharp
change on the curve of the average least square error versus
the values of IC [l]