Department of Informatics, Nicolaus Copernicus University, Grudziadzka 5, 87-100 Torun, Poland.
Computational systems are still far behind biological systems in object recognition, reasoning or analysis of language structures. What kind of data structures can be learned from data with existing machine learning algorithms? Neurocognitive inspirations show why existing learning systems cannot compete with biological ones. They point the way to more efficient algorithms, generating simplest reliable models of data and capable of object recognition with undetermined number of features. The goal of learning in neural networks and other systems is to transform data into linearly separable data clusters. This is sufficient for relatively simple problems, but makes learning almost impossible if the logic inherent in data is complex. New non-separable targets for learning are introduced to simplify learning and to characterize non-separable problems into classes of growing complexity. Neurobiological and formal justification for new learning targets are given and the case of Boolean functions analyzed.
Preprint for comments in PDF, 96 KB.
Reference: W. Duch, Learning data structures with inherent complex logic: neurocognitive perspective, The 6th WSEAS International Conference on Computational Intelligence, Man-Machine Systems and Cybernetics (CIMMACS '07), Tenerife, Canary Islands, Spain, Dec. 14-16, 2007, pp. 294-303.
BACK to the publications of W. Duch.
BACK to the on-line publications of the Department of Informatics, NCU.