The effect of noise on an optical machine learning model

Abstract

To this day machine learning has been executed by computers. In this thesis, we research the possibility, and difficulties, of an experiment that performs machine learning tasks using light propagation. This would be more energy-efficient as well as faster than current machines. We study the effect that noise, which is unavoidable in the real setup, will have on the experiment. We find that small quantities of noise during training can decrease the validation accuracy of a model with a few percents. For a model that is trained without noise, the validation effciency is not affected strongly by noise.