Two issues are of great importance for the development of real-life applications based on artificial neural networks. On the one hand, it is necessary to improve the VLSI implementation techniques, in order to increase the computational power that can be integrated on a single die. On the other, there is the need to develop design methods which take into account the constraints dictated by the hardware realizations. Concerning this topic, many researchers have focused their attention on design methodologies of continuous- and discrete-time cellular neural networks (CNNs and DTCNNs, respectively) for application to pattern recognition, associative memories and image processing [l--3]. As CNNs and DTCNNs are non-linear dynamical systems, it is necessary to analyse their stability properties before designing these networks to perform any task. To this purpose, some studies have been devoted to find the conditions which assure the asymptotic stability of the equilibrium points, as well as those which assure that each network trajectory converges to an equilibrium point. In these cases, the input data are fed via initial conditions and the outputs reach their steady-state values at an equilibrium point which depends on the initial conditions. Concerning the utilization of CNNs and DTCNNs for information storage and retrieval, it should be noted that all the methods developed until now are based on this theoretical approach [4--6]. In fact, starting from initial conditions which represent a corrupted or incompletely specified set of data, the network output is able to reconstruct exact and complete information. However, this approach presents a drawback from the VLSI implementation point of view. Namely, the initial conditions are required to be set to zero each time the network is run. Obviously, this is an undesirable feature for networks running in real time [7].
展开▼
