Self-Organizing Maps (SOM), sometimes called Kohonen networks, are a special class of neural networks. A self-organizing map consists of neurons placed at the nodes of a two-dimensional lattice. The neurons become selectively activated to various input mass spectra or classes of spectra in the course of a competitive learning process. The neurons compete among themselves to be activated or excluded. SOM can be considered as a nonlinear generalization of PCA.

The principal goal of self-organizing maps is to transform a set of n-dimensional input spectra into a discrete two-dimensional map, and to display this transformation in a user-readable fashion. Each input spectrum presented to the network activates a neuron according to a complex set of interrelationships between spectra. In SOM each mass spectrum must always activate a neuron and this spectrum is shown on the particular neuron. Spectra that activate the same neuron belong, in terms of classification, to the same pattern. To ensure that the self-organizing process has a chance to develop properly, the networks should be exposed to a certain number of different spectra. Therefore, in Mass Frontier a minimum of 10 spectra must be used in a self-organizing process.

SOM architecture used in Mass Frontier. Each spectrum Activates a Neuron in the map.

Note: The SOM classification method exhibits one atypical feature you should be aware of. Different results are produced for an identical data set if the input data is processed in a different order. This order sensibility is an inherent feature of neural networks and NOT a result of faulty algorithms.