Topics in Blind Signal Processing by Neural Networks

Abstract

The aim of this paper is to propose a brief summary of the research topics and results presented in the Author's PhD Thesis, concerning blind signal processing by artificial neural networks. The results are illustrated by the citation of the formal instruments involved in the research activity and by several links to the main published contributions. Introd uctionSeveral signal processing problems are currently tackled by means of linear and non-linear discrete-time adaptive circuits, which are capable of self-designing in order to achieve a pre-defined target.Such structures have raised an ever increasing interest in the international scientific community as they allow for solving a number of problems which turn out to be incompatible with the classical solutions; in fact, they allow to easily manage the uncertainty inherent in any design activity. Furthermore, they permit to achieve the desired results even in the partial, or sometimes total, lack of fundamental information about the statistical and temporal features of the signals under analysis, about the transmission means that the signals propagate within, and about the receivers and measurement devices used as sensors and transducers.Owing to their noticeable flexibility and complexity and their relative novelty, such adaptive structures, linear as well as non-linear, are not endowed with a complete and consistent theory which would allow their synthesis and analysis under general conditions: for this reason they appear in the literature with configurations still not completely satisfactory about their capability of achieving the desired targets, and rather inefficient about the computational and structural complexity required in order to achieve these targets. It is thus