Conditional convergence of photorefractive perceptron learning

Optoelectronic implementation of a 256-channel sonar adaptive-array processor

We present an optoelectronic implementation of an adaptive-array processor that is capable of performing beam forming and jammer nulling in signals of wide fractional bandwidth that are detected by an array of arbitrary topology. The optical system makes use of a two-dimensional scrolling spatial light modulator to represent an array of input signals in 256 tapped delay lines, two acousto-optic modulators for modulating the feedback error signal, and a photorefractive crystal for representing the adaptive weights as holographic gratings. Gradient-descent learning is used to dynamically adapt the holographic weights to optimally form multiple beams and to null out multiple interference sources, either in the near field or in the far field. Space-integration followed by differential heterodyne detection is used for generating the system's output. The processor is analyzed to show the effects of exponential weight decay on the optimum solution and on the convergence conditions. Several experimental results are presented that validate the system's capacity for broadband beam forming and jammer nulling for linear and circular arrays.

Real-time signal classification with an acousto-optic triple-product processor cascaded into a volume holographic classifier

An all-optical generalized linear machine applied to a high-bandwidth temporal signal-classification problem is demonstrated. The classifier consists of a dimensional increasing acousto-optic triple-product processor feature extractor cascaded through an optically addressed spatial light modulator into a volume holographic implementation of a linear classifier. Multiple-expo sure implementations of learning are used to train the classifier interconnection weights in a photorefractive crystal for a training set of wide-bandwidth temporal signals input to the acousto-optic triple-product processor. Experimental implementation of high-speed, time-shift and Doppler invariant, wide-bandwidth signal identification is demonstrated.

Gain and exposure scheduling to compensate for photorefractive neural-network weight decay

A gain and exposure schedule that theoretically eliminates the effect of photorefractive weight decay for the general class of outer-product neural-network learning algorithms (e.g., backpropagation, Widrow-Hoff, perceptron) is presented. This schedule compensates for photorefractive diffraction-efficiency decay by iteratively increasing the spatial-light-modulator transfer function gain and decreasing the weight-update exposure time. Simulation results for the scheduling procedure, as applied to backpropagation learning for the exclusive-OR problem, show improved learning performance compared with results for networks trained without scheduling.