4. Results

4.2 Maps

From the above datasets, the dimensionality reduction and the representation of the landscape have been made. The self-organizing map program package [Kohonen et al., 1995] has been improved in various areas and used to train the neural network and construct the unified matrix. Figure 13 shows the maps resulting from completion of these tasks. In the maps, small crosses represent locations and grayscale the landscape.

The self-organizing maps used are composed of an input layer of a number of units equal to the number of resources. The output is made over a grid lattice with 64 by 64 units.

The training of the self-organizing maps was the most computational intensive task. For the large datasets, it takes more than 24 hours of computation to complete. Hopefully, a parallel implementation of the learning algorithm is possible (see section 5.7 "Parallel implementation" on page 34).

Having completed the self-organizing maps training process, each resource was presented to the network, which responded with a winning neuron, the location of the resources on the map.

To allow interaction with these maps, they have been made available on-line on the World-Wide Web. Using HTML forms [Berner-Lee et al., 1995b] generated by programs that comply with the CGI (Common Gateway Interface) [Robinson, 1995] specifications [Grobe, 1995], a graphical user interface has been created. Therefore, the graphical user interface is made using dynamic hypermedia documents. Figure 14 depicts its visual aspect.

Cyberspace geography visualization - 15 October 1995

Luc Girardin, The Graduate Institute of International Studies