Generalized net model of temporal learning algorithm for artificial neural networks

This paper introduces a new learning algorithm based on the temporal history of the connection weights changes. The basic idea is to investigate the weight alternation frequencies in order to discriminate stable areas from unstable ones. Once determined stable areas can be replaced with topologically simpler neural structures. Unstable areas can be extended with additional neurons or can be functionally modified by changing activation and total input formation functions of the examined neurons.