Comparison of both models

Hierarchical feature maps have two benefits over self-organizing maps which make this model particularly attractive in an information retrieval setting as described in the remainder of this paper.

First, hierarchical feature maps have substantially shorter training times than self-organizing maps. The reason for that is twofold. On the one hand, there is the obvious input vector dimension reduction on the transition of one layer to the next. Shorter input vectors lead directly to reduced training times because of faster winner selection and weight vector adaptation. On the other hand, and even more important, the self-organizing training process is performed faster because the spatial relation of different areas of the input space is maintained by means of the network architecture rather than by means of the training process. A more detailed treatment of this issue may be found in [7].

Second, hierarchical feature maps may be used to produce isolated, i.e. disjoint, clusters of the input data. Moreover, these disjoint clusters are gradually refined when moving down along the hierarchy. Contrary to that, in its basic form, the self-organizing map cannot be used to produce isolated clusters. The separation of data items rather is a tricky task that requires some insight into the structure of the input data. What one gets, however, from a self-organizing map is a sort of overall representation of input data similarities. In this sense we may use the following picture to contrast the two models of neural networks. Self-organizing maps can be used to produce maps of the input data whereas hierarchical feature maps produce an atlas of the input data. Taking up this metaphor the difference between both models is quite obvious. Self-organizing maps, in our point of view, provide the user with a single picture of the underlying data archive. As long as the map is not too large, this picture may be sufficient. As the maps grow larger, however, they have the tendency of providing too little orientation for the user. In such a case we would advise to change to hierarchical feature maps as the model for representing the contents of the data archive. In this case, the data is organized hierarchically which facilitates browsing into relevant portions of the data archive. In much the same way as one would probably not use the map of the world in order to find one's way from Schönbrunn to Neustift one would probably not use a single map of a document archive to find a particular document. Conversely, when given an atlas one might follow the hierarchy of maps along a path such as World $\rightarrow$ Europe $\rightarrow$ Austria $\rightarrow$ Vienna in order to finally find the way from Schönbrunn to Neustift. In a similar way an atlas of a document archive might be used.