Recent research [Li et al., 1995, Racz and Klotz 1991, Butchart et. al. 1997, Adams et.
al. 1998, Song and Lee, 1998] has investigated the possibility that the nodes could be
arranged in a tree structure; such an arrangement has two potential benefits: searching a
tree is fast and any hierarchical information in the data may be explicitly represented in
the tree. Most clustering algorithms, neural net or otherwise, expect the number of
classes used to be predefined; this is an obvious problem if no a-priori knowledge about
the data is available. Consequently further research has been directed towards
exploring the possibility that a network can dynamically evolve the appropriate structure
in response to the data. Such an approach naturally leads to tree structures - a unit may
produce child nodes if it feels it is not classifying its local data with sufficient
granularity.