-Means Clustering Analysis

Another way of coming up with inputs sets that share a common property is using k-means clustering algorithms. This method is for gathering input patterns that produce similar activations in all of the hidden units [Medler98]. Like the bands, this will be a different output range in each of the hidden units. However it is not possible to make classification of ranges manually, especially when there are a large number of hidden units. Therefore an iterative algorithm is used to come up with input sets [Bishop95, pp. 187-189] (see Section 4.1.12).