Neural Network Modelling: Structural Analysis via Singular Value Decomposition
Guillermo B. Sentoni, Osvaldo Enrique Agamennoni, Alfredo C. Desages, JA Romagnoli · Process Industries Power the Pacific Rim: Sixth Conference of the Asia Pacific Confederation of Chemical Engineering; Twenty-first Australasian Chemical Engineering Conference; Official Proceedings of Combined Conference 1993 · 1993
The Neural Network (NN) users have to deal with two main problems: data pre-processing and NN topology specification. The data set may contain noisy, redundant and contradictory information that make difficult the neural network approximation process. With respect to the definition of the NN topology, the technique primarily utilised to date is to specify the inputs through engineering process knowledge and the hidden layer topology through 'trial and error'. In this paper we present a technique via singular value decomposition to perform a structural analysis of the data to evaluate the NN inputs as well as the number of hidden layers. Besides, a singular value decomposition based technique is also used to improve the convergence of the back propagation training algorithm.