Drift-tolerant signal processing for incoherent optical neural networks

Abstract

Incoherent optical neural network implementations using only positive light intensities require a coding method to implement bipolar signals. The actual coding method significantly influences the manufacturability, performance, and reliability of the optical neural network. This paper describes new coding methods and compares them with currently used methods. Special attention is paid to the actual hardware implementation and the overall neural network performance under the influence of drift and manufacturing tolerances. New spatial light modulator architectures enable neural network implementations that have significantly reduced sensitivity to backlight non- uniformity, sensor array non-uniformity, and tolerances and drift of LC components. Simulations show that the new coding methods reduce the sensitivity of liquid crystal light modulator-based neural network nonlinearities by more than 50%, significantly simplifying practical implementation of large neural networks.