There are clear requirements for neural processors, ranging from pattern recognition to industrial process control. It is recognised that the degree of interconnection required for efficient Neural Networks is sufficiently high that optical or hybrid (optically-interconnected electronics) implementations should have advantages over all-electronic machines. The objective of the project is targeted at the study, design and implementation of a photonic demonstrator neural network.
Our proposal falls broadly in three areas. The first concerns 'smart pixels' as potentially suitable devices for use as optoelectronic neural units. A second study involves the properties of photorefractive crystals as reconfigurable interconnections, while the third area concerns the design of new algorithms (and architectures) which best exploit the optical response characteristics of the chosen devices rather than attempting to use devices in some 'standard' algorithm and architecture combination. We believe that the strong interaction of these three areas provides the key to successful neural net construction.