Wavefront matching method as a deep neural network and mutual use of their techniques

Abstract This paper investigates correspondence between the wavefront matching method (WFM), a design method for optical circuits, and the residual network (ResNet), one of the deep neural networks (DNNs). “Lesion studies” show that there is a correspondence not only in optimization or learning but also in characteristics. We show that the skip connection, an important concept in ResNets, needs to be modified for the WFM, and propose a new optical skip connection. Considering the WFM as a DNN, we apply the dropout technique, which is common in DNNs, with the optical skip connection. To show the possibility that the physical interpretations and various techniques can be mutually used to develop the WFM and DNNs , we present, as an example, the successful reduction of the over-adaptation in the WFM.