Neural self-organization using graphs

Abstract

The self-organizing feature map (SOFM) algorithm can be generalized, if the regular neuron grid is replaced by an undirected graph. The training rule is furthermore very simple: after a competition step, the weights of the winner neuron and its neighborhood must be updated. The update is based on the generalized adjacency of the initial graph. This feature is invariant during the training; therefore its derivation can be achieved in the preprocessing. The newly developed self-organizing neuron graph (SONG) algorithm is applied in function approximation, character fitting and satellite image analysis. The results have proven the efficiency of the algorithm.