Optical Implementation of a Neural Network for Pattern Recognition

HIS REPORT DESCRIBES the construction of a dynamic optical hybrid system for implementing multi-layer neural networks. The communication between neurons is performed by amplitude modulating optical signals with dynamic transmission filters realized with a ferroelectric liquid crystal spatial light modulator (FLC-SLM). A large part of the information processing is thus performed in parallel. The amplitude modulated signals are detected by a CCD-camera and some further processing is done in a conventional computer. The system should recognize two-dimensional graphic patterns and it has been tested on the ten Arabic digits in different shapes. As neural net algorithm a modified version of the Neocognitron model of Kunihiko Fukushima has been used. The system has been simulated in MATLAB and its ability to generalize and its sensitivity to disturbances have been examined. Furthermore the possibility of using a binary FLC-SLM to perform multi-level amplitude modulation has been verified. After training on a small number of different series of the ten digits, the simulated network has capability to generalize to shapes that are not part of the training set. Unfortunately the synaptic dimensions of the network are so large that the optical implementation could not be performed with the equipment presently at our disposal. With further refined optical components this hybrid system will probably be highly competitive with systems using entirely digital computation. T