结合最大化均值差异和生成式对抗网络：无监督域自适应与图像生成

讲座题目：结合最大化均值差异和生成式对抗网络：无监督域自适应与图像生成

讲座人：左旺孟  哈尔滨工业大学计算机学院教授、博士生导师

讲座时间：9:30-10:30

讲座日期：2017年4月7日

地点：长安校区 文津楼三段5层522学术讨论室

主办单位：ABG欧博网平台登录  智能视觉计算研究团队

讲座内容：

概率分布差异程度的度量是非监督领域自适应和图像生成领域的一个核心问题，目前主要有最大化均值差异（MMD）和生成式对抗网络（GAN）两种非参数度量方式。相对而言，MMD在领域自适应中得到了较多的关注和应用，而GAN更多地被应用于图像生成。因此，我们对最近MMD和GAN方面的进展做了简单的总结。针对领域自适应，我们分析了当前MMD模型的问题和不足，提出了一种加权MMD模型，并针对非监督领域自适应情形下给出了一种权重的自适应估计方法，以及在领域自适应和图像生成任务中验证了WMMD的有效性。结合人脸属性转换问题，通过设计恰当的感知损失和正则化项，我们给出了一种基于GAN的解决方案，并进一步提出了一种自适应GAN模型，实验结果验证了自适应GAN相对于GAN在训练效率和生成效果上的优势。

讲座人简介：

左旺孟，哈尔滨工业大学计算机学院教授、博士生导师。主要从事图像增强与复原、距离度量学习、目标跟踪、图像与视频分类等方面的研究。在CVPR/ICCV/ECCV等顶级会议和T-PAMI、IJCV及IEEE Trans.等期刊上发表论文50余篇。

Connect MMD with GAN: Unsupervised Domain Adaptation and Image Generation

Title of Lecture: Connect MMD with GAN: Unsupervised Domain Adaptation and Image Generation

Lecturer: Prof. WangMeng Zuo

Time: 9:30-10:30

Date: 2017-4-7

Venue: 522 Academic Hall, School of Computer Science, Chang’an Campus

Hosted by: School of Computer Science

About the Lecture:

The probability distribution of the degree of difference measurement is a core problem of unsupervised domain adaptation and image generation field, there are differences between the maximum mean (MMD) and generative network (GAN) against two types of nonparametric measure. Relatively speaking, MMD has received more attention and application in the field of adaptive, and GAN has been applied to image generation. Therefore, we made a brief summary of recent MMD and GAN developments. In the field of adaptive, we analyzed the current problems of MMD model and shortcomings, this paper proposes a weighted MMD model, and the supervision of the field of adaptive adaptive estimation method under given a weight, as well as in the field of adaptive and image generation tasks to verify the effectiveness of WMMD. With the face attribute conversion, through the design of appropriate perception loss and regularization, we propose a solution based on GAN, and proposes an adaptive GAN model, experimental results show that the adaptive GAN relative to GAN in the training efficiency and the effect of the advantage.

Profile of the Lecturer:

Zuo WangMeng, Professor, doctoral tutor, School of computer science, Harbin Institute of Technology. His interest is image enhancement and restoration, distance metric learning, target tracking, image and video classification. He has published more than 50 papers in CVPR/ICCV/ECCV and other top conferences and T-PAMI, IJCV and IEEE Trans. and other journals.