Título: A Backpropagation with Automatically Generated Momentum Method
Autores: Rios Neto, Atair; Vogler, Osmar
Resumo: A method is developed for the iterative parallel solution of the feedforward neural networks supervised training problem. Stochastic optimal linear estimation is used go get a backpropagation with momentum method, where the momentum weighting is automatically done. The estimation of neural network weights is done avoiding the difficulties characteristic of Kalman filtering type of algorithms and related with the adjustment and calculations involving the a priori covariance matrix of estimation errors. Preliminary numerical testing indicates that the method is a competitive choice in terms of effectiveness, efficiency and facility of use when compared to the backpropagation method.
Código DOI: 10.21528/CBRN2003-048
Artigo em PDF: 6CBRN_048.PDF
Arquivo BibTex: 6CBRN_048.bib