Self-organising neural networks: their application to "real-world" problems

This paper discusses the use of a variety of self organising neural networks to solve industrial and other "real-world" problems. In particular an application within the oil industry dealing with the classification of fluorescence spectrograms from crude oil samples and another application involving image processing in which microscopic charcoal and pollen are classified so that they can be used to estimate the contribution of tropical land clearance to global climate change. Both tasks are currently performed manually but results are time consuming to obtain and are often inaccurate. Three self organising neural networks are considered, the Self-organising Feature Map, an ART network and a self-training, deformable template matching neural network called PARADISE. The operational characteristics of the networks are discussed with particular reference to the "real-world" applications.