2007 | OriginalPaper | Buchkapitel
ChapBoltzmann Machines Learning Using High Order Decimation
verfasst von : Enric Farguell, Ferran Mazzanti, Eduardo Gomez-Ramirez
Erschienen in: Hybrid Intelligent Systems
Verlag: Springer Berlin Heidelberg
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Boltzmann Machines are recurrent and stochastic neural networks that can learn and reproduce probability distributions. This feature has a serious drawback in the exhaustive computational cost involved. In this context, decimation was introduced as a way to overcome this problem, as it builds a smaller network that is able to reproduce exactly the quantities required to update the weights during learning. Decimation techniques developed can only be used in sparsely connected Boltzmann Machines with stringent constraints on the connections between the units. In this work, decimation is extended to any Boltzmann Machine with no restrictions on connections or topology. This is achieved introducing high order weights, which incorporate additional degrees of freedom.