A cascaded recurrent neural network for real-time nonlinear adaptive filtering

A new form of recurrent neural network, referred to as a cascaded recurrent neural network (CRNN), is described. This network can perform temporally extended tasks. A learning procedure is described for adjusting the weights in the network in order to produce a desired input-output relation in the time domain. An important feature of CRNNs is that they can perform real-time nonlinear adaptive filtering. This application is illustrated by exploring the nonlinear prediction of chaotic signals.<>