Dynamical functional artificial neural networks (D-FANN) for intelligent signal processing

For the sake of clarity and precision, intelligent signal processing will be defined as the process of mapping a signal x into a binary vector y, so that y enables the detection, classification, or interpretation of an event present in x. We denote by f the input-output map of such an intelligent signal processing filter. In a number of applications it is possible to implement the nonlinear filter map f by an artificial neural network (ANN). In the present paper we consider the case in which the signal x is an analog signal (waveform) belonging to L ²(I), where I is an appropriate interval of the real line R ¹, and propose the realization of f by an artificial neural network in which the synaptic weight actions of the first layer are implemented by a filter bank. We call such a network a dynamical functional artificial neural network (D-FANN) to distinguish it from a conventional functional artificial neural network (FANN), where a synaptic weight action is implemented by a scalar product in L² (I), between the incoming waveform x and a functional weight. Thus, compared with conventional FANNs, D-FANNs permit simple and meaningful causal realizations of intelligent analog signal processors. A novel element is the introduction of a D-FANN gain equation, in a way analogous to that in Kalman Filtering. Applications of D-FANNs to real and simulated data are now in progress and these results will be presented at the symposium (ISCAS-98)