Kohonen Networks on Transputers: Implementation and Animation

Abstract

Self organizing feature maps have been introduced in 1982 by the Finish phycicist Kohonen [KOHO 82]. Since then they have been used for a variety of applications. Implementations of the algorithm on conventional hardware are rather slow for big problems; direct VLSI or special purpose hardware implementations are rather expensive.

In this paper we describe an implementation of the algorithm on a network of transputer. The network makes efficient use of the algorithm’s inherent parallelism. The computational power of the net can easily be extended to almost any desired range by adding more processors; the ratio of price to performance is very good as only off-the-shelf components are used. The implementation allows flexible reconfiguration and adaption to all network and vector sizes.

The network offers a speed of up to 2.7 Mega CUPS. This allows to train even fairly big nets of more than 10,000 units within less than 30 minutes. These good performance characteristics give the possibility to animate the training process in real time. The resulting pictures are not only aesthetic in their own right but give some insight into the algorithm’s behaviour at the same time.