Interpolating self-organising map (iSOM

A new learning algorithm is presented for enhancing the scale or structure of an already trained self-organising map (SOM) without the need to re-use the original training data. Alternative methods for the insertion of these additional interpolating neurons, while still preserving the learnt topology, are presented together with two illustrative examples of the algorithm in operation. Introduction: The self-organising map (SOM) [2, 31 is an unsuper-vised learning algorithm that clusters and projects potentially high dimensional input data onto a discrete neural grid or map of usu-ally reduced dimensions. It is often applied with a predetermined size and structure as the correct size is rarely known a priori. Sev