基于Winslow泛函的生成模型
发布时间:2022-07-20 19:09
概率生成模型,也叫作生成模型,是在机器学习和概率统计问题中的一类具有极高实际应用价值的模型。它的应用十分广泛,可以用来对不同种类的数据进行建模,比如图像,声音,文本数据,同时它能够通过多种方式融入强化学习,所以在数据预测,图片处理,文本生成等领域有广泛的作用。但是如何设计一种高效且有效的生成模型,也是非常具有挑战性的。生成模型的关键步骤就是对目标分布进行参数化估计。为了在一定程度上简化讨论,在本文中我们将主要关注通过极小化交叉熵(KL)的原理工作的生成模型。生成模型的种类非常多,但是主要能分为两类,一类是构造一个显式的密度分布。在这些显式的密度模型中,密度是可以计算处理的,所以模型的更新也是相对直接的。比如变分自编码器。另一类生成模型没有显式地表示数据所在空间上的概率分布,相反,该模型提供了某种方式来减少与这种概率分布的直接交互。通常是直接提取样本的能力,比如使用马尔科夫链来随机变换现有样本的方法,以便从同一分布中获得另一个样本。特别的,有一类特别的具有显式密度函数的生成模型,是基于定义两个不同空间之间的连续非线性变换来构造的,称为流模型。换句话说,这类模型从一个简单的分布出发,将其与...
【文章页数】:71 页
【学位级别】:硕士
【文章目录】:
摘要
Abstract
Chapter 1 Introduction
1.1 Source, background, and significance
1.2 Research Status
1.2.1 Worldwide Research Status
1.2.2 Domestic Research Status
1.3 Thesis Outline
Chapter 2 Generative model
2.1 Preliminaries and Notations
2.2 Minimizing KL divergence through flow maps
2.3 Stein variational gradient descent
2.4 Summary
Chapter 3 Adaptive gird method
3.1 Grid distribution based on the equidistribution principle
3.2 Grid distribution based on the variational principle
3.3 Iterative grid redistribution
3.4 Summary
Chapter 4 Generative model driven by Winslow functional
4.1 The choice of monitor function
4.2 Generative model based on Winslow map
4.2.1 Theoretical results
4.2.2 Algorithm
4.2.3 Numerical example
4.3 Generative model based on mapping G
4.3.1 Theoretical results
4.3.2 Algorithm
4.3.3 Numerical example
4.4 Summary
Chapter 5 Generative model implementation by DNN
5.1 Preliminaries
5.1.1 The Deep Ritz Method
5.1.2 Generative adversarial network
5.2 Architecture of neural network
5.3 Numerical examples
5.4 Summary
Conclusions
结论
References
Appendix
Acknowledgements
本文编号:3664592
【文章页数】:71 页
【学位级别】:硕士
【文章目录】:
摘要
Abstract
Chapter 1 Introduction
1.1 Source, background, and significance
1.2 Research Status
1.2.1 Worldwide Research Status
1.2.2 Domestic Research Status
1.3 Thesis Outline
Chapter 2 Generative model
2.1 Preliminaries and Notations
2.2 Minimizing KL divergence through flow maps
2.3 Stein variational gradient descent
2.4 Summary
Chapter 3 Adaptive gird method
3.1 Grid distribution based on the equidistribution principle
3.2 Grid distribution based on the variational principle
3.3 Iterative grid redistribution
3.4 Summary
Chapter 4 Generative model driven by Winslow functional
4.1 The choice of monitor function
4.2 Generative model based on Winslow map
4.2.1 Theoretical results
4.2.2 Algorithm
4.2.3 Numerical example
4.3 Generative model based on mapping G
4.3.1 Theoretical results
4.3.2 Algorithm
4.3.3 Numerical example
4.4 Summary
Chapter 5 Generative model implementation by DNN
5.1 Preliminaries
5.1.1 The Deep Ritz Method
5.1.2 Generative adversarial network
5.2 Architecture of neural network
5.3 Numerical examples
5.4 Summary
Conclusions
结论
References
Appendix
Acknowledgements
本文编号:3664592
本文链接:https://www.wllwen.com/kejilunwen/yysx/3664592.html