【作者】 张琪；

【导师】 林一青；

【作者基本信息】 上海交通大学 ， 应用统计（专业学位）， 2020， 硕士

【摘要】 生成对抗网络(GANs)是一种深度神经网络架构,是人工智能领域的研究热点。随着深度学习的发展,研究者们越来越关注生成式模型的研究。生成式模型对真实世界进行建模运算,学习数据内部的统计规律并生成类似的样本。这一过程涉及大量的先验知识和庞大的计算量,因此相较于判别式模型,生成式模型的发展相对缓慢。生成式对抗网络的提出,为生成式模型的研究提供了新的思路,在学术界和工业界都获得了广泛的关注。WGANs是对原始GANs的经典改进。WGANs提出一个全新的损失函数,从本质上解决了对生成分布和真实分布的衡量问题,使得GANs的训练有一个指示指标,且从理论上规避了梯度消失导致的训练失败的问题。RWGANs是在WGANs的基础上,进一步一般化损失函数,放宽WGANs中的代价函数必须是对称函数的要求,为分布距离的衡量问题提供了有价值的补充。本文在RWGANs的基础上,对其损失函数进行进一步的研究讨论,提出了一个新的下界函数去近似原函数,并在此基础上提出了新的RWGANs算法。新算法解决了原算法中共轭网络的训练问题,且与WGANs具有类似的形式,可以看作在WGANs的基础上引入函数变换,且不同的函数选择会对训练产生不同的影响。本文通过MNIST、Fashion-MNIST和CIFAR数据集上的对比实验,验证了我们的猜想,并得出了新算法能够在不降低图片质量的情况下使GANs的训练更加稳定,收敛更快的结论。更多还原

【Abstract】 Generative Adversarial Nets(GANs)are a kind of deep neural network architectures,which is a research hotspot in the field of artificial intelligence.Along with the development of deep learning,researchers pay more atten-tion to the research of generative models.The generative models simulate the real world,learn the statistical patterns from data and generate similar samples.This process involves a lot of prior knowledge and a large amount of computation,so the development of generative models is relatively slow,compared with discriminative models.GANs provide a new thought for the research on generative models,which have gained extensive attention from both academia and industry.WGANs improve original GANs by introducing a new loss function,which effectively measuring the discrepancy on the distribution of the gen-erated and real data.This function is an indicator of the GANs training effi-ciency,and helps to avoid the problem of training failure caused by gradient disappearance.On the basis of WGANs,RWGANs further extend the loss function and relax the requirement that the cost function must be symmetric in WGANs.In the framework of RWGANs,we propose a new lower bound function to approximate the original loss function and suggests a new RWGANs algo-rithm.The new algorithm solve the training problem of conjugate network in the original RWGANs algorithm,and have a similar form with WGANs.It can be regarded as a function transformation before the training of WGANs,and different choices of functions impact on the efficiency of training.In this paper,comparison experiments on MNIST,Fashion-MNIST and CI-FAR data sets are conducted to verify our hypothesis.In summary,the new algorithm can make the GANs training be more stable and converge faster without reducing the image quality.更多还原