PUB - Publikationen an der Universität Bielefeld

Neural networks have emerged as an efficient method to complement more traditional approaches, in particular in situations where a design of algorithms from first principles becomes too costly or fails due to insufficient information about e.g. the statistics of a problem. However, as the problems to which neural networks are applied become more demanding, such as in machine vision, the choice of an adequate network architecture becomes more and more a crucial issue. This is particularly true for larger applications, where the actions of several neural networks need to be coherently integrated into a larger system. Unfortunately, systematic investigations of this issue are just beginning to appear in the literature (for an interesting approach, see e.g. [2, 12]) and results are still rather sparse.