At one time, analogue and digitial VLSI techniques were seen as equally valid approaches to neural network solutions. This is no longer the case. Digitial techniques have proved more appropriate for conventional neural algorithms, and artificial nerual networks (ANNs) have themselves developed in directions that make hardware implementations much more difficult. We describe our experiences of implemening different kinds of ANNs (MLP, RBF, Hebbian) and draw a number of lessons from this experience. We then indicate the circumstances in which aanlogue VLSI techniques might prove mst beneficial in the future, using three exemplar applications to illustrate the point,
展开▼
