Optical adaptive neural networks with a ground glass for global random interconnections and local gain controls

TAG (training by adaptive gain) is a new adaptive neural network model for large-scale optical implementations. For single-layer neural networks withN input andM output neurons, the TAG model contains two different types of interconnections, i.e.MN global fixed interconnections and (N+M) adaptive gain controls. The training algorithm is based on gradient descent and error back-propagation, and is easily extensible to multilayer and/or higher-order architectures. For large-scale electrooptic implementation, the fixed global interconnections may be implemented by multifacet hologram, volume hologram or ground glass, and the adaptive gains by spatial light modulators (SLMs). The ground glass is more advantageous for random interconnections, with much higher diffraction efficiency and interconnection density. Both feed-forward signal and error back-propagation paths are implemented by a single ground glass, and adaptive learning has been demonstrated for heteroassociative memory and classifier.