Novel training algorithm for limited connected networks

J. C. Wang, R. O. Grondin

    Research output: Chapter in Book/Report/Conference proceedingConference contribution

    Abstract

    We have successfully applied this algorithm to several simple problems including some which are nonlinearly separable. An important advantage of this training algorithm is that we do not need to relearn the whole set of learning vectors if new data are added to the system. This is a significant improvement over the back propagation algorithm where new data requires relearning the whole training set. No algorithm however can teach a network functions that the network is incapable of performing. We have discovered that, contrary to the hope of one of the present authors that one could trade extra layers for connectivity, there exist problems which cannot be done on a limited fan-in system.

    Original languageEnglish (US)
    Title of host publicationIEE Conference Publication
    PublisherPubl by IEE
    Pages387-389
    Number of pages3
    Edition313
    StatePublished - 1989
    EventFirst IEE International Conference on Artificial Neural Networks - London, Engl
    Duration: Oct 16 1989Oct 18 1989

    Other

    OtherFirst IEE International Conference on Artificial Neural Networks
    CityLondon, Engl
    Period10/16/8910/18/89

    ASJC Scopus subject areas

    • Electrical and Electronic Engineering

    Fingerprint

    Dive into the research topics of 'Novel training algorithm for limited connected networks'. Together they form a unique fingerprint.

    Cite this