Kohonen networks

The second neural networks paradigm that is implemented in the library is the Kohonen Topology Preserving Mapping Algorithm [Kohonen 1989]. The following sets of routines are implemented:

Network creation and manipulation

Kohonen networks with intrinsic dimension ranging from one to four can be generated and manipulated. The functions used for creation and manipulation are koh_draw_init, make_matrix_3D, koh_weight_draw and koh_draw_free.

Network training

A Kohonen network can be trained with the standard Kohonen learning algorithm and some of its variants. Among these are, the algorithm that only updates the units in a small and limited neighborhood around the unit that is closest to the data input vector, and the algorithm that updates all units, but weighed with a function of the distance to the closest unit, e.g. see [Kohonen 1989]. The functions used in this context are koh_adapt_unit, koh_adapt_net and koh_learn.