These graphics, which have a pattern like a colourful chessboard and also appear elsewhere in the exhibition, depict self-organising maps known as Kohonen maps. They map an artificial neural network. Each small square field represents a neuron – the map is therefore like a virtual “cross-section of the brain”. The dots scattered across the square surface represent data that the map sorts (“maps”) according to their similarity to each other.
The map learns through repeated entry of data. At first, the neurons are in a random state, processing the data whose properties are most similar to their own. They then gradually adapt their own state to that of the data. This eventually lends the neural network its structure, as it “learns” to distinguish the properties of the entered data better and better. In the application phase, the trained map can then sort (“map”) the data to be analysed.
The properties according to which the entered data are to be compared are determined by the programming. The results are therefore comprehensible. For research purposes, Kohonen maps have a clear advantage over connectionist models. The latter work with network connections and it remains unclear how they arrive at their results.
Now try your hand at training a Kohonen map in the interactive station “Learning colours”!