Coherent correlator design analysis for the implementation of deep learning networks

Abstract

Optical signal processing can reduce the electrical power consumption required over that required by graphically processing units. There remain a number of challenges to overcome. Noise is potentially much larger in optical systems than in their electronic counterparts, and some of the operations, such as a bias addition, are not easy to implement in free space processors. This paper analyzes a proposed design that utilizes a camera and lightweight electronic processing to perform the convolutional layers. Simulations are performed to compare the expected performance against an ideal system that cannot be physically realized and a proposed architecture. The impact of speckle noise in the system is analyzed and methods to reduce this are proposed.