On the accuracy of mapping by neural networks trained by backpropagation with forgetting

Mapping properties of multi-layer, feedforward artificial neural networks are analyzed using modified backpropagation training with forgetting (decay) of the connection weights. Neural nets trained by forgetting algorithm are not sensitive to the initial choice of the network, and the trained network structure can be used for knowledge acquisition regarding the feature classes. The accuracy of the non-linear mapping realized by layered neural networks is limited in the sense of minimum classification error and it can be estimated based on the a posteriori probability densities of the training classes. It is shown in this paper that backpropagation with forgetting is a convenient tool to implement finite accuracy of learning. The proposed strategy has been used for anomaly detection in actual time series. It is shown that neural networks trained by forgetting algorithm have better generalization capabilities than those trained by standard backpropagation. The analyzed feature classes have been characterized by making use of the information extracted from the structure of the trained network.