@TechReport{TR031,
  author =	 "Elmar Thurner",
  institution =	 "PRIP, TU Wien",
  number =	 "PRIP-TR-031",
  title =	 "Modulare {N}eurale {S}systeme, {A}ufgabenstellung der {D}iplomarbeit",
  year =	 "1994",
  url =		 "https://www.prip.tuwien.ac.at/pripfiles/trs/tr31.pdf",
  abstract =	 "During the last years neural networks have been
                  applied to a wide range of pattern recognition
                  applications. The reasons for this rapid growth of
                  interest for neural networks are the innovative
                  capabilities of neural networks: Neural networks are
                  universal and adaptive function approximators. They
                  provide self organization as well as distributed
                  representation of knowledge. Despite of these
                  advantages, the solutions found, suffered by new
                  problems like the enormous effort in computational
                  power to train the networks, bad scaling properties
                  and insufficient extendability. In this report it is
                  stated, that the reason for these disadvantages can
                  be found in the architecture of the used
                  networks. It will be shown, that a single general
                  purpose network will not fit optimally to any given
                  problem. The report gives an introduction to a new
                  approach to overcome the mentioned problems:
                  Modularity. Starting with basic definitions like
                  modular and hierarchical system, a detailed analysis
                  of the advantages of modular architectures in
                  comparison to large, single networks is
                  presented. Using these definitions the principal
                  structures, found in modular neural architectures,
                  like parallelism (integrative and competitive),
                  cascades and supervisor actor structures are
                  described. Finally an outlook to the goals of the
                  diploma theses and a definition of performance
                  criteria for neural networks for optical character
                  recognition are given.",
}