前两天纽约暴雪,天地一片苍茫。今天元宵节,长岛依然清冷寂寥,正月十五闹花灯的喧嚣热闹已成为悠远的回忆。这学期,老顾在讲授一门研究生水平的数字几何课程,目前讲到了2016年和丘成桐先生、罗锋教授共同完成的一个几何定理【3】,这个工作给出了经典亚历山大定理(Alexandrov Theorem)的构造性证明,也给出了最优传输理论(Optimal Mass Transportation)的一个几何解释。这几天,机器学习领域的Wasserstein GAN突然变得火热,其中关键的概念可以完全用我们的理论来给出几何解释,这允许我们在一定程度上亲眼“看穿”传统机器学习中的“黑箱”。下面是老顾下周一授课的讲稿。
为此,我们引入最优传输的几何理论(Optimal Mass Transportation),这个理论可视化了W-GAN的关键概念,例如概率分布,概率生成模型(生成器),Wasserstein距离。更为重要的,这套理论中,所有的概念,原理都是透明的。例如,对于概率生成模型,理论上我们可以用最优传输的框架取代深度神经网络来构造生成器,从而使得黑箱透明。
总结
通过以上讨论,我们看到给定两个概率分布This is the rendered form of the equation. You can not edit this directly. Right click will give you the option to save the image, and in most browsers you can drag the image onto your desktop or another program.,则存在唯一的一个凸函数(Brenier 势函数)This is the rendered form of the equation. You can not edit this directly. Right click will give you the option to save the image, and in most browsers you can drag the image onto your desktop or another program.,其梯度映射This is the rendered form of the equation. You can not edit this directly. Right click will give you the option to save the image, and in most browsers you can drag the image onto your desktop or another program.把一个概率分布This is the rendered form of the equation. You can not edit this directly. Right click will give you the option to save the image, and in most browsers you can drag the image onto your desktop or another program.映成了另外一个概率分布。这个最优传输映射的传输代价就给出了两个概率分布之间的Wasserstein距离。Brenier势能函数,Wasserstein距离都有明晰的几何解释。
参考资料
[1]Arjovsky, M. & Bottou, L.eon (2017) Towards Principled Methods for Training Generative Adversarial Networks
[2] Arjovsky, M., Soumith, C. & Bottou, L.eon (2017) Wasserstein GAN.
[3] Xianfeng Gu, Feng Luo, Jian Sun and Shing-Tung Yau, Variational Principles forMinkowski Type Problems, Discrete Optimal Transport, and Discrete Monge-Ampere
Equations, Vol. 20, No. 2, pp. 383-398, Asian Journal of Mathematics (AJM), April 2016.