PARALLEL NONLINEAR ADAPTIVE DIGITAL-FILTERS USING RECURRENT NEURAL NETWORKS

In signal processing applications, it sometimes happens that a problem that is difficult to process based on the linear theory can be solved by nonlinear processing. When the unknown system is nonlinear, howeve r, it is difficult to process the estimation problem in real time by t he conventional nonlinear processing method, since a large amount of c omputation is required in order to determine the optimal solution. On the other hand, neural networks with a nonlinear input-output relation are considered as interesting and are applied to various problems suc h as pattern recognition and the estimation of nonlinear systems. This paper considers a nonlinear adaptive digital filter that has a large number of parameters, with a requirement for real-time processing A me thod of designing a parallel recurrent neural digital filter by introd ucing multiple, small-scale recurrent neural networks is proposed. Com paring the proposed method to the conventional method based on the lin ear or nonlinear theories, a better result is obtained by the proposed method. As another aspect, learning efficiency can be improved by the proposed method since parallel learning is executed.