Modulare neuronale Systeme (bibtex)
by Elmar Thurner
Abstract:
This thesis gives an introduction and an overview to a new approach in the field of neural networks: Modularity. It will be shown, that a single general purpose network will not fit optimally to any given problem. Contrary, modular networks learn faster, have better generalization abilities and higher robustness and are easier to extend. To investigate the relationship between architecture and function the basic structures, found in modular neural architectures, like parallelism (integrative and competitive), cascades and supervisor actor structures are explored. Furthermore algorithms are presented to estimate the quality of results of networks, to combine the results of net- works, to automatically decompose tasks and to combine different architectures. The theoretical analysed advantages of modular neural networks are demonstrated by experiments in the field of optical character recognition, OCR. For the recognition of printed characters in 20 different fonts by using modularity the training time is reduced to less than a quarter and the miss classification rate by one third in compari- son to the single network solution. For the classification of hand written digits using modular neural preprocessing, an improvement of the miss classification rate of 7\% is achieved compared to the non modular network.
Reference:
Modulare neuronale Systeme (Elmar Thurner), Technical report, PRIP, TU Wien, 1995.
Bibtex Entry:
@TechReport{TR038,
  author =	 "Elmar Thurner",
  institution =	 "PRIP, TU Wien",
  number =	 "PRIP-TR-038",
  title =	 "Modulare neuronale {S}ysteme",
  year =	 "1995",
  url =		 "ftp://ftp.prip.tuwien.ac.at/pub/publications/trs/tr38.ps.gz",
  abstract =	 "This thesis gives an introduction and an overview to
                  a new approach in the field of neural networks:
                  Modularity. It will be shown, that a single general
                  purpose network will not fit optimally to any given
                  problem. Contrary, modular networks learn faster,
                  have better generalization abilities and higher
                  robustness and are easier to extend. To investigate
                  the relationship between architecture and function
                  the basic structures, found in modular neural
                  architectures, like parallelism (integrative and
                  competitive), cascades and supervisor actor
                  structures are explored. Furthermore algorithms are
                  presented to estimate the quality of results of
                  networks, to combine the results of net- works, to
                  automatically decompose tasks and to combine
                  different architectures. The theoretical analysed
                  advantages of modular neural networks are
                  demonstrated by experiments in the field of optical
                  character recognition, OCR. For the recognition of
                  printed characters in 20 different fonts by using
                  modularity the training time is reduced to less than
                  a quarter and the miss classification rate by one
                  third in compari- son to the single network
                  solution. For the classification of hand written
                  digits using modular neural preprocessing, an
                  improvement of the miss classification rate of 7\%
                  is achieved compared to the non modular network.",
}
Powered by bibtexbrowser