Abstract: Neural networks with delay elements and feedback can be powerful tools for signal processing applications. Classical design of systems for signal processing is based on three basic assumptions: linearity, stationarity and the existence of second order statistics (usually Gaussian). These basic assumptions are required for the statistical and mathematical development but they are not usually satisfied on real signals and systems even if in some cases they can be approximately true. In the real world the results of a classical signal processing filter design may be not optimal. By means I will show, artificial neural networks can be used successfully with classical signal processing techniques to improve the performance of the systems. I will give a review of dynamic neural network architectures valuable for signal processing, system identification-control, signal classification or in general for temporal processing. I will explain some new results on how these architectures can be applied to real time temporal processing. In addition, a new general learning paradigm inspired by Tellegen's Theorem, will be presented The new theorem can be applied to any non-linear circuit such as a dynamic neural network.
Back to the LEMS homepage.