



In optical multiplexing, the orthogonality between channels plays a crucial role. This orthogonality ensures that signals between different channels do not interfere with each other, enabling efficient data transmission. The optical multiplexing system can transmit multiple channels of data at the same time, effectively improving the utilization of optical fibers. However, such a system also inevitably imposes an upper limit on multiplexing capacity.
Here, the Key Laboratory of Synaesthetic Fusion Photonics Technology of the Ministry of Education of Guangdong University of Technology developed a non-orthogonal optical multiplexing on multi-mode fiber (MMF) based on a deep neural network, Called Speckle light field retrieval network (SLRnet), it can learn complex mapping relationships between multiple non-orthogonal input light fields containing information encoding and their corresponding single intensity outputs.
Through principle verification experiments, SLRnet successfully solved the ill-posed problem of non-orthogonal optical multiplexing on MMF. It is able to utilize a single shot speckle output to unambiguously retrieve multiple non-orthogonal input signals mediated by the same polarization, wavelength and spatial location with 98% fidelity. This research paves the way for realizing high-capacity optical multiplexing utilizing non-orthogonal channels and is an important step towards this goal.
This research will promote potential applications in the fields of optics and photonics and provide new insights into the exploration of broader disciplines such as information science and technology.
The related research was titled "Non-orthogonal optical multiplexing empowered by deep learning" and was published in "Nature Communications" on February 21, 2024.
Optical multiplexing problem
Multiplexing is the cornerstone of optical communications, in which physical orthogonality between multiplexed channels is a prerequisite for large-scale coded information transmission.
Considering the demultiplexing of multiple orthogonal signals, transmission matrix methods (such as MMF) can even solve this problem on strongly scattering media.
Recently, deep learning has been widely used in the fields of optics and photonics for the reverse design of optical devices and computational optics. Specifically, deep neural networks have been used to improve the performance of orthogonal multiplexing over multiple scattering media.
However, all reported multiplexing scenarios to date strictly rely on physical orthogonality between multiplexed channels. There have been no attempts to leverage the nonlinear modeling capabilities of deep learning to achieve non-orthogonal optical multiplexing over MMFs.
Unfortunately, the multiplexing of non-orthogonal channels mediated by the same polarization or wavelength even in single-mode fiber is still very challenging due to the lack of efficient demultiplexing methods or digital signals Processing is overburdened. Therefore, developing a new method to decode the information encoded in non-orthogonal input channels is critical for eventual optical multiplexing.
Non-orthogonal optical multiplexing on MMF based on deep neural network
Here, researchers demonstrate that preliminary non-orthogonal optical multiplexing can be achieved via MMF with the support of SLRnet.
As a proof-of-concept demonstration, non-orthogonal input channels can be used to achieve multiplexed transmission of information through MMF, including general natural scene images, unrelated random binary data and images that do not belong to the same type of training data set, It is beneficial to realize non-orthogonal multiplexing transmission of optical information.
Through data-driven techniques to establish complex relationships between non-orthogonal input channels and outputs, trained deep neural networks can retrieve non-orthogonal Encoded information for traffic lanes. Even non-orthogonal multiplexed channels sharing the same polarization, wavelength and input spatial region can be efficiently decoded.
Neural Network Architecture
DepthNeural NetworkAble to retrieve non-orthogonal optical multiplexing from a single speckle output of an MMF Signal. Multiple amplitude- and phase-encoded information mediated by arbitrary polarization combinations can be efficiently retrieved by SLRnet after propagating in the MMF.
As shown in Figure 2a, even the typical scenario of non-orthogonal input channels with the same polarization, wavelength and input space region can be explicitly decoded. This is achieved through a deep neural network, the architecture of which is shown in Figure 2b, which is a variant of Unet based on the unique multiple scattering process of MMF. It consists of a fully connected (FC) layer and ResUnet.
Experimental results
First consider the case where the MMF length is 1m. Figure 3a shows the evolution of the retrieval fidelity for two multiplexed light field channels with arbitrary combinations of polarization states during SLRnet training. Overall, there will be four encoding channels in the amplitude and phase dimensions, which can be non-orthogonal depending on the polarization state. Retrieval fidelity was measured by Pearson's correlation coefficient (PCC).
As can be seen from the figure, the evolution of PCC retrieved using the same SLRnet training configuration after 100 epochs is greater than 0.97. At the same time, the evolution of retrieval fidelity is essentially the same for the twelve multiplexing scenarios, demonstrating the excellent robustness of non-orthogonal multiplexing to arbitrary polarization combinations.
Additionally, Figure 3b provides the fidelity retrieved in each amplitude and phase multiplexed channel separately using different polarization combinations. The average retrieval fidelity in the amplitude and phase dimensions is nearly identical (~0.98), highlighting the ability of SLRnet to demultiplex information encoded in multiple non-orthogonal input channels.
In order to perform sensory evaluation of the retrieval information of wavefront encoding, the typical demultiplexing results of four polarization combinations (0° and 0°, 0° and 10°, 0° and 90°, and 0° and ellipse) are as follows: Figure 4 shows.
It can be seen that four grayscale images multiplexed on the amplitude and phase of the input wavefront using the same polarization can be effectively demultiplexed using a single speckle output. Other results retrieved with similar fidelity under different polarization combinations demonstrate that SLRnet is capable of unprecedented non-orthogonal input channel multiplexing even when the encoding wavefront is scrambled by MMF.
To further consolidate the superiority of SLRnet in more realistic scenarios, non-orthogonal optical multiplexing results using the same polarization state on 50 m MMF are proposed, as shown in Figure 5. As can be seen from Figures 4 and 5, the demultiplexing results for the 1 m MMF are better than the 50 m case because the scattering properties of the longer MMF are more susceptible to environmental influences. Demultiplexing performance can be further improved by optimizing the network structure. Research shows that SLRnet is an effective means of multiplexing non-orthogonal channels in MMF.
Finally, to demonstrate the versatility of SLRnet for different image sets, the study shows that SLRnet has good generalization.
Although the MMF-based non-orthogonal optical multiplexing concept proposed at this stage cannot be directly used for medical diagnosis that usually requires uniform fidelity, high-precision non-orthogonal multiplexing of non-correlated binary digital information It shows that the realization of non-orthogonal multiplexing transmission of optical information through MMF is a step forward.
This research may not only pave the way for utilizing high-throughput MMFs for communication and information processing, but may also provide a paradigm shift for optical multiplexing in optics and other fields, which can greatly increase the degree of freedom and freedom of optical systems. capacity.
The above is the detailed content of The fidelity is as high as ~98%. The "AI + Optics" research of Guangzhou University of Technology is published in the Nature sub-journal. Deep learning empowers non-orthogonal optical multiplexing.. For more information, please follow other related articles on the PHP Chinese website!

人工智能Artificial Intelligence(AI)、机器学习Machine Learning(ML)和深度学习Deep Learning(DL)通常可以互换使用。但是,它们并不完全相同。人工智能是最广泛的概念,它赋予机器模仿人类行为的能力。机器学习是将人工智能应用到系统或机器中,帮助其自我学习和不断改进。最后,深度学习使用复杂的算法和深度神经网络来重复训练特定的模型或模式。让我们看看每个术语的演变和历程,以更好地理解人工智能、机器学习和深度学习实际指的是什么。人工智能自过去 70 多

众所周知,在处理深度学习和神经网络任务时,最好使用GPU而不是CPU来处理,因为在神经网络方面,即使是一个比较低端的GPU,性能也会胜过CPU。深度学习是一个对计算有着大量需求的领域,从一定程度上来说,GPU的选择将从根本上决定深度学习的体验。但问题来了,如何选购合适的GPU也是件头疼烧脑的事。怎么避免踩雷,如何做出性价比高的选择?曾经拿到过斯坦福、UCL、CMU、NYU、UW 博士 offer、目前在华盛顿大学读博的知名评测博主Tim Dettmers就针对深度学习领域需要怎样的GPU,结合自

一. 背景介绍在字节跳动,基于深度学习的应用遍地开花,工程师关注模型效果的同时也需要关注线上服务一致性和性能,早期这通常需要算法专家和工程专家分工合作并紧密配合来完成,这种模式存在比较高的 diff 排查验证等成本。随着 PyTorch/TensorFlow 框架的流行,深度学习模型训练和在线推理完成了统一,开发者仅需要关注具体算法逻辑,调用框架的 Python API 完成训练验证过程即可,之后模型可以很方便的序列化导出,并由统一的高性能 C++ 引擎完成推理工作。提升了开发者训练到部署的体验

深度学习 (DL) 已成为计算机科学中最具影响力的领域之一,直接影响着当今人类生活和社会。与历史上所有其他技术创新一样,深度学习也被用于一些违法的行为。Deepfakes 就是这样一种深度学习应用,在过去的几年里已经进行了数百项研究,发明和优化各种使用 AI 的 Deepfake 检测,本文主要就是讨论如何对 Deepfake 进行检测。为了应对Deepfake,已经开发出了深度学习方法以及机器学习(非深度学习)方法来检测 。深度学习模型需要考虑大量参数,因此需要大量数据来训练此类模型。这正是

Part 01 概述 在实时音视频通信场景,麦克风采集用户语音的同时会采集大量环境噪声,传统降噪算法仅对平稳噪声(如电扇风声、白噪声、电路底噪等)有一定效果,对非平稳的瞬态噪声(如餐厅嘈杂噪声、地铁环境噪声、家庭厨房噪声等)降噪效果较差,严重影响用户的通话体验。针对泛家庭、办公等复杂场景中的上百种非平稳噪声问题,融合通信系统部生态赋能团队自主研发基于GRU模型的AI音频降噪技术,并通过算法和工程优化,将降噪模型尺寸从2.4MB压缩至82KB,运行内存降低约65%;计算复杂度从约186Mflop

导读深度学习已在面向自然语言处理等领域的实际业务场景中广泛落地,对它的推理性能优化成为了部署环节中重要的一环。推理性能的提升:一方面,可以充分发挥部署硬件的能力,降低用户响应时间,同时节省成本;另一方面,可以在保持响应时间不变的前提下,使用结构更为复杂的深度学习模型,进而提升业务精度指标。本文针对地址标准化服务中的深度学习模型开展了推理性能优化工作。通过高性能算子、量化、编译优化等优化手段,在精度指标不降低的前提下,AI模型的模型端到端推理速度最高可获得了4.11倍的提升。1. 模型推理性能优化

今天的主角,是一对AI界相爱相杀的老冤家:Yann LeCun和Gary Marcus在正式讲述这一次的「新仇」之前,我们先来回顾一下,两位大神的「旧恨」。LeCun与Marcus之争Facebook首席人工智能科学家和纽约大学教授,2018年图灵奖(Turing Award)得主杨立昆(Yann LeCun)在NOEMA杂志发表文章,回应此前Gary Marcus对AI与深度学习的评论。此前,Marcus在杂志Nautilus中发文,称深度学习已经「无法前进」Marcus此人,属于是看热闹的不

过去十年是深度学习的“黄金十年”,它彻底改变了人类的工作和娱乐方式,并且广泛应用到医疗、教育、产品设计等各行各业,而这一切离不开计算硬件的进步,特别是GPU的革新。 深度学习技术的成功实现取决于三大要素:第一是算法。20世纪80年代甚至更早就提出了大多数深度学习算法如深度神经网络、卷积神经网络、反向传播算法和随机梯度下降等。 第二是数据集。训练神经网络的数据集必须足够大,才能使神经网络的性能优于其他技术。直至21世纪初,诸如Pascal和ImageNet等大数据集才得以现世。 第三是硬件。只有


Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

AI Hentai Generator
Generate AI Hentai for free.

Hot Article

Hot Tools

Dreamweaver Mac version
Visual web development tools

MinGW - Minimalist GNU for Windows
This project is in the process of being migrated to osdn.net/projects/mingw, you can continue to follow us there. MinGW: A native Windows port of the GNU Compiler Collection (GCC), freely distributable import libraries and header files for building native Windows applications; includes extensions to the MSVC runtime to support C99 functionality. All MinGW software can run on 64-bit Windows platforms.

MantisBT
Mantis is an easy-to-deploy web-based defect tracking tool designed to aid in product defect tracking. It requires PHP, MySQL and a web server. Check out our demo and hosting services.

Atom editor mac version download
The most popular open source editor

Notepad++7.3.1
Easy-to-use and free code editor
