光华讲坛——社会名流与企业家论坛第5287期

主题：Deep Generative Learning via Variational Gradient Flow

主讲人：香港科技大学 杨灿博士

主持人：统计学院 林华珍教授

时间：2019年4月18日（星期四）下午3:00-4:00

地点：西南财经大学柳林校区弘远楼408会议室

主办单位：统计研究中心 统计学院 科研处

主讲人简介：

杨灿，香港科技大学助理教授，健康数据分析实验室主任。从事统计，生物统计，以及机器学习方面的研究。在包括Annals of Statistics, PNAS, Nature Communication,The American Journal of Human Genetics, Bioinformatics等在内的国际期刊发表论文70多篇。截至2019年3月，杨博士的作品被引用2,383次，其中h指数为22，i10指数为35。

主要内容：

Learning the generative model, i.e., the underlying data generating distribution, based on large amounts of data is one of the fundamental tasks in machine learning and statistics. Recent progresses in deep generative models have provided novel techniques for unsupervised and semi-supervised learning, with broad application varying from image synthesis, semantic image editing, image-to-image translation to low-level image processing. However, statistical understanding of deep generative models is still lacking, e.g., why the logD trick works well in training generative adversarial network (GAN). In this talk, we introduce a general framework, variational gradient flow (VGrow), to learn a deep generative model to sample from the target distribution via combing the strengths of variational gradient flow on probability space, particle optimization and deep neural network. The proposed framework is applied to minimize the f-divergence between the evolving distribution and the target distribution. We prove that the particles driven by VGrow are guaranteed to converge to the target distribution asymptotically. Connections of our proposed VGrow method with other popular methods, such as VAE, GAN and flow-based methods, have been established in this framework, gaining new insights of deep generative learning. We also evaluated several commonly used f-divergences, including Kullback-Leibler, Jensen-Shannon, Jeffrey divergences as well as our newly discovered “logD” divergence which serves as the objective function of the logD-trick GAN. Experimental results on benchmark datasets demonstrate that VGrow can generate high-fidelity images in a stable and efficient manner, achieving competitive performance with state-of-the-art GANs.

This is a joint work with Yuan Gao, Yuling Jiao, Yao Wang, Yang Wang and Shunkang Zhang.