生成对抗网络GAN综述

生成对抗网络(GAN)作为一种新的无监督学习算法框架得到越来越多研究者的青睐,已然成为当下的一个研究热点。GAN受启发于博弈论中的二人零和博弈理论,其独特的对抗训练思想能生成高质量的样本,具有比传统机器学习算法更加强大的特征学习和特征表达能力。目前GAN在计算机视觉领域尤其是在样本生成领域取得显著成功,每年有大量GAN相关研究的论文产出。针对GAN这一热点模型,首先介绍了GAN的研究现状;接着介绍了GAN的理论、框架,详细分析了GAN在训练过程中存在梯度消失和模式崩溃的原因;然后讨论了一些典型的GAN的改进模型,总结了它们理论的改进之处、优点、局限性、应用场景以及实现成本,同时还将GAN与VAE、RBM模型进行比较,总结出GAN的优势和劣势;最后展示了GAN在数据生成、图像超分辨率、图像风格转换等方面的应用成果,并探讨了GAN目前面临的挑战以及未来的研究方向。     

Effective Routing and Scheduling in Adversarial Queueing Networks

In an adversarial queueing network, the incoming traffic is decided by an adversary, who operates under a reasonable rate restriction. This model provides a valuable, complementary point of view to that of the traditional queueing network model in which arrivals are modeled by stochastic processes. As a result, the adversarial queueing network model has attracted much attention in recent years, especially as a way of modeling packet injections into a communication network. Our main result is a simple, effective packet routing and scheduling algorithm with a provably good performance. Specifically, our algorithm keeps the system stable (bounded number of packets in the system), with the bound on the number of packets in the system that is O((1 - r)^-1), where r can be interpreted as the arrival rate of the packets into the communication network. This improves upon the work of Gamarnik, who designed an algorithm for which the number of packets in the system is O((1 - r)^-2); moreover, our result matches the traditional queueing-theoretic number-in-system bound.     

战术地域分组网

本文概述外军战术地域通信网络和交换技术,介绍叠加在电路交换方式上的PTARMIGAN和MSE战术分组网络     

Loss-Sensitive Generative Adversarial Networks on Lipschitz Densities

International Journal of Computer Vision - In this paper, we present the Lipschitz regularization theory and algorithms for a novel Loss-Sensitive Generative Adversarial Network (LS-GAN)....     

生成式对抗网络在图像补全中的应用*

图像补全是数字图像处理领域的重要研究方向,具有广阔的应用前景。提出了一种基于生成式对抗网络(GAN)的图像补全方法。生成式对抗网络模型由生成器模型和判别器模型两部分构成,通过采用卷积神经网络(CNN)实现。首先,通过生成器模型对图像的缺失区域进行补全;然后,利用判别器模型对图像的补全效果进行判别。采用马尔科夫随机场(MRF)与均方误差(MSE)相结合的损失函数训练生成器模型,加强对图像纹理细节的处理能力。实验结果证明,基于生成式对抗网络的图像补全方法,相较于其他现有的方法,具有更好的补全效果。     

多尺度生成式对抗网络图像修复算法

图像修复作为深度学习领域的一个研究热点,在人们现实生活中有着重要的意义。现有图像修复算法存在各种问题,导致视觉上无法达到人们的要求。针对现有图像修复算法精确度低、视觉一致性差以及训练不稳定等缺陷,提出了一种基于生成式对抗网络(GAN)模型的图像修复算法。该算法主要对判别器的网络结构进行改进,在全局判别器和局部判别器的基础上引入多尺度判别器。多尺度判别器在不同分辨率的图像上进行训练,不同尺度的判别器具有不同的感受野,能够引导生成器生成更全局的图像视图以及更精细的细节。针对GAN训练中经常出现的梯度消失或梯度爆炸问题,使用WGAN(Wasserstein GAN)的思想,采用EM距离模拟样本数据分布。在CelebA、ImageNet以及Place2图像数据集上对该算法的网络模型进行了训练和测试,结果显示与先前的算法模型相比,该算法提高了图像修复的精确度,能够生成更为逼真的修复图片,并且适用于多种类型图片的修复。     

基于生成式对抗网络的链路预测方法

近些年来,网络中链路预测问题逐渐兴起。相比于传统启发性模型,以神经网络为基础的链路预测方法由于其能够自我学习的优点,逐渐获得研究者的青睐。结合生成式对抗网络,一种创新性的链路预测方法WL-GAN(Weisfeiler-Lehman generative adversarial networks)被提出。WL-GAN首先利用子图提取算法与子图编码算法,为网络中的每条已知关系的节点对构造以该节点对为结构中心的节点对子图,并获得相应连接矩阵。随后,利用矩阵数据来训练生成式对抗网络,最终可以获得能够判断子图中心节点对是否存在链路的判别器。实验结果表明,WL-GAN拥有优秀的性能与稳定性。     

On the Accuracy of Packet Delay Estimation in Distributed Service Networks

Packet delay (either one-way time or round-trip time) is a very important metric for measuring the performance of networks in a highly dynamic environment such as the Internet. Many network applications are also sensitive to packet delay or delay variation for ensuring an acceptable level of quality in providing network-based services such as VoIP, multimedia streaming, etc. A very important property of packet delay is that it is very dynamic and therefore should be measured frequently with measurement results being updated on a timely basis. Measurement of packet delay has thus generated a great deal of interest in the past years and a lot of research has been performed in the development of measurement architecture as well as specific measurement techniques. However, how to reduce network overhead resulting from measurement while achieving a reasonable level of accuracy still remains a challenge. In this paper, we propose to use delay estimation as an alternative to delay measurement for reducing measurement overhead and, in particular, examine the level of accuracy that delay estimation can achieve. With delay estimation, measurement nodes can be dynamically selected and activated and other nodes can share measurement results by performing delay estimation, thus reducing measurement overhead while supporting the dynamic requirement for delay measurement. Consequently, while measurement overhead can be reduced by activating only a subset of network nodes to perform actual measurement, desired accuracy can be achieved by exploring the correlation between delays as well as by sharing measurement results to do delay estimation based on such a correlation. We illustrate how packet delays of network nodes can correlate to each other based on topological properties and show how delays can be estimated based on such a correlation to meet accuracy requirements, which would make delay measurement in the Internet highly dynamic and adaptable to the accuracy requirements and measurement results highly reliable. We also show how delay estimation can be applied by presenting three application scenarios as well as an example to demonstrate the usefulness and effectiveness of delay estimation in the measurement of packet delays.     

基于生成式对抗网络的图像修复研究进展

图像修复是图像处理的一个重要问题,目的是利用计算机视觉技术自动恢复退化图像中损坏或丢失的部分,被广泛应用于影视特技制作、图像编辑、数字化文物保护等领域。近几年,以生成式对抗网络(GAN)为代表的深度学习技术在计算机视觉和图像处理领域大获成功,基于GAN的图像修复逐渐成为主流,受到了广泛关注。针对图像修复的关键问题,文章对GAN和基于GAN的修复方法进行理论分析,首先整理分析了传统的基于人工特征的经典图像修复方法,其次总结了近年来基于GAN的代表性图像修复算法,并进行归纳分类,探讨了各类方法的特点和局限性。然后对图像修复模型常用的评价指标和公开数据集进行整理和分析,最后阐述了图像修复面临的挑战,对图像修复技术未来的发展方向进行展望。     

基于循环神经网络和生成式对抗网络的口令猜测模型研究

深度学习技术的进展为提高口令猜测效率提供了潜在的新途径.目前,已有研究将循环神经网络(Recurrent Neural Network,RNN)、生成式对抗网络(Generative Adversarial Network,GAN)等深度学习模型运用于设计口令猜测模型.本文基于RNN模型、概率上下文无关文法(Proba...     

生成对抗神经网络在心电异常识别中的应用研究

针地在使用深度学习方法构建心电异常识别模型,常常由于心电异常事件样本分布不平衡,造成机器识别心电异常事件模型性能表现差的问题,论文提出一种基于生成对抗网络的数据增强方法来获取均匀分布的训练数据集,其主要过程如下:首先使用小波变换消除心电信号中噪声,然后使用压缩感知模型压缩心电信号来减少网络中的参数,利用生成对抗神经网络模型扩充数据集,最后使用卷积神经网络建立分类模型.实验结果表明,使用对抗神经网络能够显著改善数据集中样本分布不均衡的问题,平均F1达到了98.73％.引入压缩感知模型后,在不影响模型性能表现的情况下,将模型参数量减少了28.30％.基于对抗神经网络的数据增强方法可以有效地解决心电异常分类过程样本分布不均衡,利用压缩感知模型方法不仅可以保证模型性能,同时降低了模型的复杂程度.     

基于生成对抗网络的语音信号分离

基于深度学习的单声道语音分离需要计算时频掩蔽,但现有语音分离方法中时频掩蔽不可学习,也未将其封装到深度学习中进行优化,通常依赖于维纳滤波法进行后续处理。为此,提出一种基于生成对抗网络的语音信号分离方法。在语音生成阶段引入递归推导算法和稀疏编码器来改进时频掩蔽生成结果,并将生成的语音输入至判别器进行分类,以降低信号源之间的扰动。实验结果表明,与基于深度神经网络的语音信号分离方法相比,该方法的SDR、SIR分离指标分别提高6.2 dB和5.0 dB。     

生成式对抗网络研究与应用进展

基于零和博弈思想的生成式对抗网络(GAN)可通过无监督学习获得数据的分布,并生成较逼真的数据。基于GAN的基础概念及理论框架,研究各类GAN模型及其在特定领域的应用情况,从数据相似性度量、模型框架、训练方法3个方面进行分析,对GAN改进与扩展的相关研究成果进行总结,并从图像合成、风格迁移等应用领域展开讨论,归纳出GAN的优势与不足,同时对其应用前景进行展望。分析结果表明,GAN的学习能力与可塑性强,改进潜力大,应用范围广,但其发展面临的挑战是训练过程不稳定,且缺乏生成数据质量的客观评价标准。     

Performance of Scheduling Policies in Adversarial Networks with Non-synchronized Clocks

In this paper we generalize the Continuous Adversarial Queuing Theory (CAQT) model (Blesa et al. in MFCS, Lecture Notes in Computer Science, vol. 3618, pp. 144–155, 2005) by considering the possibility that the router clocks in the network are not synchronized. We name the new model Non Synchronized CAQT (NSCAQT). Clearly, this new extension to the model only affects those scheduling policies that use some form of timing. In a first approach we consider the case in which although not synchronized, all clocks run at the same speed, maintaining constant differences. In this case we show that all universally stable policies in CAQT that use the injection time and the remaining path to schedule packets remain universally stable. These policies include, for instance, Shortest in System (SIS) and Longest in System (LIS). Then, we study the case in which clock differences can vary over time, but the maximum difference is bounded. In this model we show the universal stability of two families of policies related to SIS and LIS respectively (the priority of a packet in these policies depends on the arrival time and a function of the path traversed). The bounds we obtain in this case depend on the maximum difference between clocks. This is a necessary requirement, since we also show that LIS is not universally stable in systems without bounded clock difference. We then present a new policy that we call Longest in Queues (LIQ), which gives priority to the packet that has been waiting the longest in edge queues. This policy is universally stable and, if clocks maintain constant differences, the bounds we prove do not depend on them. To finish, we provide with simulation results that compare the behavior of some of these policies in a network with stochastic injection of packets.     

使用孪生注意力机制的生成对抗网络的研究

生成对抗网络(GAN)能够生成逼真的图像,已成为生成模型中的一个研究热点。针对生成对抗网络无法有效提取图像局部与全局特征间依赖关系以及各类别间的依赖关系,提出一种用于生成对抗网络的孪生注意力模型(TAGAN)。以孪生注意力机制为驱动,通过模拟局部与全局特征间的依赖关系以及各类别间依赖关系,对真实自然图像建模,创建逼真的非真实图像。孪生注意力机制包含特征注意力模型和通道注意力模型,特征注意力模型通过有选择地聚合特征,学习相似特征间的关联性,通道注意力模型通过整合各通道维度的相关特征,学习各通道的内部依赖关系。在MNIST、CIFAR10和CelebA64数据集上验证了所提出模型的有效性。     

Non-expected Route Choice Model under Risk on Stochastic Traffic Networks

In this paper, we propose a novel non-expected route travel time (NERTT) model, which belong to the rank-dependent expected utility model. The NERTT consists of two parts, which are the route travel time distribution and the distortion function. With the strictly increasing and strictly concave distortion function, we can prove that the route travel time in the proposed model is risk-averse, which is the main focus of this paper. We show two different reduction methods from the NERTT model to the travel time budget model and mean-excess travel time model. One method is based on the properly selected distortion functions and the other one is based on a general distortion function. Besides, the behavioral inconsistency of the expected utility model in the route choice can be overcome with the proposed model. The NERTT model can also be generalized to the non-expected disutility (NED) model, and some relationship between the NED model and the route choice model based on the cumulative prospect theory can be shown. This indicates that the proposed model has some generality. Finally, we develop a non-expected risk-averse user equilibrium model and formulate it as a variational inequality (VI) problem. A heuristic gradient projection algorithm with column generation is used to solve the VI. The proposed model and algorithm are tested on some hypothetical traffic networks and on some large-scale traffic networks.     

CubeGAN: Omnidirectional Image Synthesis Using Generative Adversarial Networks

We propose a framework to create projectively-correct and seam-free cube-map images using generative adversarial learning. Deep generation of cube-maps that contain the correct projection of the environment onto its faces is not straightforward as has been recognized in prior work. Our approach extends an existing framework, StyleGAN3, to produce cube-maps instead of planar images. In addition to reshaping the output, we include a cube-specific volumetric initialization component, a projective resampling component, and a modification of augmentation operations to the spherical domain. Our results demonstrate the network's generation capabilities trained on imagery from various 3D environments. Additionally, we show the power and quality of our GAN design in an inversion task, combined with navigation capabilities, to perform novel view synthesis.     

基于生成对抗网络的位置隐私博弈机制

文中提出了一种以用户为中心的位置隐私博弈机制,目的是在满足LBS服务质量的基础上生成对应的保护策略,并减小计算规模和效用损失.该机制以Stackelberg博弈模型为基础,用户在请求LBS服务时,采用位置模糊机制对自身位置进行扰动后发送给LBS服务器,使攻击者难以推测自己的真实位置;攻击者根据已知的一部分背景知识,对匿...     

基于深度卷积生成对抗网络的图像生成

图像生成是虚拟现实技术(virtual reality,VR)中的重要技术手段,针对传统图片生成方法需要大量的数据集进行训练,且生成的图片轮廓不清晰等问题,采用基于深度卷积神经网络和生成对抗网络来实现图片的生成.为了保证生成图片的真实性和完整性,在图片生成阶段引入变分自编码器,通过编码器获取到输入图片数据的均值和方差,...