A DCT-based D-FANN for nonlinear adaptive time series prediction

A nonlinear adaptive time-series predictor has been developed using a new type of artificial neural network called dynamical-functional artificial neural network (D-FANN) for its underlying model structure. D-FANNs are two-layer neural systems in which the synaptic weights of the first layer are “functions” rather than numbers, and where the action of a synapse on a signal passing through it takes place in the form of a scalar product in L² between the functional weight and the signal. The second layer of these networks is a combiner, which optimally linearly combines the weighted outputs of the zero-memory nonlinear elements comprising the neurons. In this brief, we introduce a neural network which we call a DCT based D-FANN. This is a D-FANN where the functional weights of the first layer are a filter bank built up of discrete cosine transform basis functions. We show that this system can successfully be used to model and predict an important class of highly dynamic and nonstationary signals, namely speech signals