Learning dynamics of photorefractive neural networks

NOTE: Text or symbols not renderable in plain ASCII are indicated by [...]. Abstract is included in .pdf document. This thesis investigates the optical implementation of neural networks utilizing dynamic photorefractive volume holography. The number of accessible degrees of freedom in a general holographic interconnection system is derived, and a cascaded-grating scheme that provides full, nondegenerate interconnections between two unsampled planes is presented. The dynamics of the formation of photorefractive volume holograms is considered. The impact of time-constant asymmetry on multiple hologram recording is evaluated. A basic framework for controlling the dynamics of photorefractive holograms is described and a number of dynamic copying methods for rejuvenating decayed holograms are identified. Experiments of linear dynamic copying using phase conjugation and nonlinear copying using an optical feedback loop are presented. The electrical fixing of photorefractive holograms in [...] crystals is experimentally demonstrated and the physical mechanism is discussed. A number of neural learning algorithms are investigated for optical implementation. An Anti-Hebbian local learning algorithm is proposed to simplify the optical architecture of feedforward multilayer networks. Experimental demonstrations of several optical neural networks are presented. An optical perceptron is trained for face classification, and the use of dynamic copying for improving its performance is demonstrated. A two-layer network based on Kanerva's sparse, distributed memory model is implemented and trained for real-time handwritten character recognition. Finally an optical two-layer network for real-time face recognition, with moderate tolerance to shift, rotation, scale, and facial expression, is presented