SGER: Investigating Optical Hardware, Fixed-Weight Learning Neural Networks

Missouri State University, Springfield MO

Investigators

Abstract

Current optical neural networks have limited learning speed. Typically, light attenuators imaged on a photographic slide control the optical couplings (weights) between the neurons. Expensive, high-speed spatial light modulators sometimes control these weights. In either case, the weights must be changed as learning takes place. Creating the new weights is slow because it does not take advantage of the optical network's speed. Developing a new slide slows some networks even more. This project will to move the learning operation from the attenuators to the light interactions themselves. Light signals (sometimes called "flying weights") encode the mapping being learned. We gain two advantages: the learning takes place at optical speeds and inexpensive slides can still be used. We move the learning to the light interactions by including fixed weight learning networks in the slide, increasing its complexity by up to an order of magnitude. However, this is usually not a problem since a slide has such high resolution. The learning really does take place at the speed of light, which is potentially one million times faster than other neural networks, either digitally-programmed or optical ones. The PI will develop an optical neural network which will solve a simple problem (performing all 16 two-input logic operations). Varied hardware and software will be investigated seeking improved speed, integration, and efficiency. These include: 1) Photodetector arrays for the output neurons, 2) High-speed interface for a micro-mirror array, 3) The use of geometric optics for neuron functions, and 4) Improved fixed-weight learning algorithms.