Currently, there is an enormous amount of interest in developing modern learning methods for massive data-driven applications—most prominently in healthcare, biological sciences, computer vision, and natural language processing. In contrast to classic learning methods such as linear regression and Principal Component Analysis (PCA), our understanding of the strengths and weaknesses of modern approaches is still developing. In particular, the robustness and scope of applicability of these algorithms remain elusive in most cases.

In this talk, I will aim to bridge the gap between theory and practice for modern learning methods by drawing principled connections with classic learning algorithms under appropriate baseline setups. I will demonstrate the success of this approach in two fundamental problems in machine learning and statistics, namely (1) learning nonlinear dependency measures among random variables and (2) learning probabilistic models from data. For the first problem, I will introduce Maximally Correlated PCA as a multivariate extension of Maximal Correlation and a nonlinear generalization of PCA. For the second problem, I will discuss Generative Adversarial Networks (GANs) by drawing connections to optimal transport theory, supervised learning and rate-distortion theory. During the talk, I will examine applications in real datasets including a cutting-edge single-cell RNA-seq dataset. 

Bio- Soheil Feizi is a post-doctoral research scholar at Stanford University in the area of machine learning and statistical inference. He received his Ph.D. in Electrical Engineering and Computer Science (EECS) with a minor degree in Mathematics from the Massachusetts Institute of Technology (MIT). He completed a M.Sc. in EECS at MIT, where he received the Ernst Guillemin award for his thesis, as well as the Jacobs Presidential Fellowship and the EECS Great Educators Fellowship. He also received the best student award at Sharif University of Technology from where he holds his B.Sc.