Fixed-weight learning neural networks on optical hardware

A. Steven Younger, MSU JVIC
Emmett Redd, Missouri State University

Abstract

Fixed-weight learning embeds a learning algorithm into the neural network topology, so its learning can take advantage of all speed increases in its operation on optical neural hardware, up to 10,000 x conventional networks. We developed a hardware-in-the-loop Optical Hardware-based Neural Network Test Apparatus. We used the apparatus to research and develop various embedded learning methods; to work out alignment, calibration, and noise reduction methods; study synaptic weight and neural signal encoding; and to test several small fixed-weight learning neural networks.