The input patterns can be grouped into sets according to their common properties looking at the internal structure of the trained network. We are interested in how the network separates inputs at the hidden units after training. Then we can analyze these sets to observe what attributes were favored to bring the inputs together, as we describe in Section 3.2.4.
The following sections summarize the methods available to find input sets that we are interested in, namely banding analysis and custering analysis.
-Means Clustering Analysis