Learning algorithms for compression and evaluation of information from large data sets

This paper deals with the problem of information compression and similarity analysis of large data sets that represent various operating conditions of machines and complex systems. Three different unsupervised learning algorithms for information compression are presented. Two of them are for off-line learning with fixed and growing number of neurons, while the third one is a special on-line learning algorithm. A comparative analysis of these algorithms reveals their merits and features. For similarity analysis the so called Key Point models are introduced. They extract the most essential features from the compressed information model as key points. A respective algorithm for similarity analysis, based on the pair wise comparison of the key points from the models is introduced. The proposed learning algorithms and the whole technology for similarity analysis are illustrated on the examples of several images.