Architectures for Optical Neural Networks

Artificial neural networks are parallel processing systems which have applications in speech and pattern recognition (Rumelhart and McCelland, 1986; Prager et al., 1986; Lippmann, 1987; Szu, 1986; Geman and Geman, 1984; Luttrell, 1985; Widrow et al. 1988), function optimization (Geman and Geman, 1984; Barhen et al., 1987; Hinton et al., 1984; Kirkpatrick et al., 1983; Hopfield and Tank, 1986), robotics (Barhen et al., 1987), and control (Psaltis et al., 1988a). They consist of a set of identical, nonlinear processing elements, generically known as neurons, which are linked together to form a highly interconnected network. Information is represented by the pattern of activity of the neurons and data are stored by distributing them throughout the network’s connections. This is done by weighting the links with positive and negative values to indicate the effect that one neuron has on another. This makes the system fault tolerant since each neuron is connected to many others and each weight represents the ‘average’ stimulus over the data set, so the loss of a few connections does not drastically affect the operation of the network.