Kohonen nets -- examples

This self teaching net has been demonstrated with some success on a number of examples.

1. An artificial example is a robot arm constrained to move in two dimensions as an ``upper arm'' and ``lower arm'' connected by an elbow joint. The position of the hand may be specified either by co-ordinates in the usual manner, or by the angles and giving the direction of the lower and upper arm respectively. Initialising a 2D neuron array at random, we might feed randomly chosen pairs to each neuron. Determining the maximum response for a given pair, weights from the 2 inputs will be updated within a shrinking rectangular neighbourhood. The network eventually learns a Cartesian representation of the input space despite being fed an equivalent but highly non-linear representation
2. A widely quoted (and more useful) example is the Kohonen Phonetic Typewriter. A 2D array of neurons, each of which have 15 inputs, is fed the Fourier coefficients of a speech signal that is sampled every 9.83msec. The spoken words are a stream of phonemes around which the samples cluster; after training in exactly the same way as the simple example above, the neurons in the network come to represent these phonemes. A test signal can then be played through the network; it will spend most of its time in the neighbourhood of the network phonemes that make up the unknown speech.

   The nature of natural language means that some phoneme sets are very hard to disambiguate, and subsidiary networks are used to analyse these. The idea has been demonstrated using the Finnish and Japanese languages.

   It must be stressed that the Kohonen network is learning these representations in an entirely unsupervised manner.