A NEW RECURRENT NEURAL NETWORK FOR TEMPORAL SIGNAL PROCESSING

摘要

In this paper, a new recurrent neural network structure, termed the Recurrent Functionally Expanded Neural Network (RFENN), is presented. The structure employed local output feedback and its learning algorithm is derived. The performance of the RFENN structure is shown to be superior to the feedforward FENN and another more complex recurrent network model for the problem of off-line modelling of the chaotic Mackey Glass Equation. In the case of real-time non-stationary speech prediction, which is fundamental to Adaptive Differential Pulse Code Modulation (ADPCM) coding-decoding, a new RFENN based RFENN-Finite Impulse Response (FIR) hybrid predictor model is shown to offer significantly superior modeling performance compared to both the linear as well as the feedforward and recurrent neural network based predictor models studied.