Optical neural networks for image analysis: imaging spectroscopy and production systems

An optical implementation of multispectral image data processing and production systems processing symbolic scene data using networks is described. A binary network is used for propositional logic and an analog network is used for reasoning with uncertainties. In the production system, the authors assign the ´trueness´ of facts to the activities of the neurons and to encoding the rules as interconnection strengths. With binary neurons, this type of network operates on a knowledge base formulated in propositional calculus to make binary-valued decisions. When analog neurons are used, reasoning with uncertainty is possible. The authors details an optical matrix-vector multiplication architecture with nonlinear elements and feedback as implementation of the system. It allows new inferences on subsequent iterations, whereas other neural systems do not allow new rules to be learned.<>