Learning from Data by Neural Networks with a Limited Complexity

0405074 - UIVT-O 20030188 RIV IT eng C - Konferenční příspěvek (zahraniční konf.)
Kůrková, Věra - Sanguineti, M.
Learning from Data by Neural Networks with a Limited Complexity.
Artificial Neural Networks in Pattern Recognition. Florence: University of Florence, 2003 - (Gori, M.; Marinai, S.), s. 146-151
[IAPR TC3 Workshop /1./. Florence (IT), 12.09.2003-13.09.2003]
Grant CEP: GA ČR GA201/02/0428
Grant ostatní: IT-CZ Area MC6(XX) Project 22
Výzkumný záměr: AV0Z1030915
Klíčová slova: learning from data * neural networks * kernel methods
Kód oboru RIV: BA - Obecná matematika

Learning from data formalized as a minimization of a relularized empirical error is studied in terms of approximate minimization over sets of functions computable by networks with increasing number of hidden units. There are derived upper bounds on speed of convergence of infima achievable over networks with n hidden inits to the global infimum. The bounds are expressed in terms of norms tailored to the type of network units and moduli of continuity of regularized empirical error functionals.
Trvalý link: http://hdl.handle.net/11104/0125289