|
With the development of artificial intelligence technology, such as artificial neural networks, the increasing demand for computing drives the upgrading of computing accelerators. It’s known that the semiconductor process is approaching physical limits and the Von Neumann architecture of storage-computing separation affects the computing efficiency, which both lead to the gradual failure of electronic devices to meet application requirements. Optical neural networks (ONNs) can take full advantage of high speed, high bandwidth, high parallelism, and low power consumption of optical transmission to overcome the deficiencies of electronic devices. In this paper, we summarize and analyze previous researches on optical neural networks according to different physical implementations. And we conclude that most studies apply the characteristics of special materials to realize the dense matrix multiplication and nonlinear activation function of ONNs. Less research focuses on the nonlinear characteristics inherent in the optical signal transmission to realize important components of traditional neural networks. ONNs show great potentials in analog computing and information processing, such as marine in-situ imaging and optical receiver of underwater optical communication. And ONN is possible to be a new generation of neural network accelerator. But the large-scale application of ONNs requires more studies in optical implementation of nonlinear activation function and loss function, and accuracy improvement of optical computing.
|
