Alternative Neural Networks and Applications
Date
2022-12
Authors
Li, Yimeng
Major Professor
Wang, Zhengdao
Advisor
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Altmetrics
Abstract
Compared to all-electronic neural networks, optical neural networks such as all-optical diffractive deep neural network proposed recently demonstrate great advantages in terms of low power consumption, light-speed computation, and zero memory consumption. However, the lack of nonlinearity in all-optical neural networks limits its performance on complex machine learning tasks. In this report, we combine optical neural networks with electronic neural network, to take advantage of power efficiency and massive parallel processing of optical computing as well as nonlinearity of electronic neural network. We demonstrate the feasibility of combining optical and electronic neural networks for the benefit of reduced power consumption and faster computation. We choose generative adversarial networks (GANs) as example applications, and show that by integrating optical and electronic networks, it is possible to design generative adversarial network that strikes tradeoff between training, reconfigurability, power consumption, and computation speed. The optical-electronic GANs (OE-GANs) can get slightly degraded performance as electronic GANs with much lower power consumption, lower memory need, and faster speed. We also apply optical-electronic networks on variational autoencoders (VAEs). The optical-electronic VAEs (OE-VAEs) can reduce the issue of mode collapse from OE-GANs, and generate some images with more variation.
Series Number
Journal Issue
Is Version Of
Versions
Series
Academic or Administrative Unit
Type
creative component
Comments
Rights Statement
Copyright
2022