Error-bound for the non-exact SVD-based complexity reduction of the generalized type hybrid neural networks with non-singleton consequents
Orsolya Takács, Annamária R. Várkonyi-Kóczy · 2002
The main advantage of neural networks (NNs) is that they are able to solve complicated problems, even if the exact mathematical model is not known. However, there is no universal method for the approximation of the proper size of the neural networks which usually results in the overestimation of the needed size. Therefore, the need arises to have formal methods for the complexity reduction of neural networks. Singular Value Decomposition (SVD) based complexity reduction was first proposed for various fuzzy inference systems. Recently, the method has been extended to generalized neural network, which made possible the use of neural networks in time-critical systems. Beyond the elimination of redundancy, the SVD-based reduction can be used to achieve further reduction, if a certain amount of error can be tolerated. This paper gives an error-bound for this further complexity reduction of generalized type hybrid neural networks with non-singleton consequents.