Figure 8 also illustrates one of the benefits of using this set of data. The visual
character of the data makes it possible to see the nature of the analysis provided by the
network. It can be seen that the weight vectors move from rough generalisations at the
top of the tree structure to more detailed pictures at the lower levels. In general the tree
splits the lighter and darker images into separate sub-trees. The left hand sub-tree has
classified the lighter and smaller images and the other two sub-trees classifying darker
and larger images with the right hand sub-tree classifying the darkest images. The next
level nodes classify according to the general orientation of the image. For example in
the left hand sub-tree the first and second classifying vertical images; the second deals
with the darker variants; the third classifies the more diagonal elements; and the fourth
classifies the horizontal images.