Parameter tolerances and generalisation abilities of cellular neural networks
O. Kufudaki, M. Novak · 2003
The problem of cellular neural network parameter tolerances is discussed with special regard to the network generalization abilities. The authors point out that from an analysis of the parameter tolerances for individual cells an estimation can be made for the whole cellular neural structure (e.g., through expansion in series of nonlinear functions in the set of given points in the training regions). In the case of cellular neural networks the expected accuracy of such an estimation can be good, because the respective nonlinear transformation is applied only once (for a one layer network) and the mathematical expressions of the expanded series are related for sparse weight matrices only.>