March 6 (Mon) @ 10:00am: "Modern Statistical and Computational Perspectives on Tensor Learning and Transfer Learning," Changxiao Cai, Postdoctoral Researcher, U. of Pennsylvania
Abstract
A fundamental task in modern data science applications is to extract relevant knowledge from large-scale data. This talk explores how to design computationally efficient and statistically optimal algorithms to learn from tensor data, and how to transfer useful information from different but related source datasets to a target domain of interest.
In the first part, I will discuss noisy tensor completion, namely the problem of reconstructing a low-rank tensor from highly incomplete and randomly corrupted observations of its entries. I'll introduce a two-stage nonconvex algorithm --- (vanilla) gradient descent following a rough initialization --- that achieves the best worlds of estimation, inference, and computational efficiency simultaneously.
In the second part, I will talk about transfer learning for contextual multi-armed bandits under the covariate shift model. I will discuss how to quantify the contribution of the dataset collected from the source domains for the target bandit learning, and introduce novel transfer learning procedures that achieve near-optimal theoretical guarantees.
Bio
Changxiao Cai is currently a postdoctoral researcher at the University of Pennsylvania. He received his Ph.D. in Electrical Engineering from Princeton University in 2021, advised by H. Vincent Poor and Yuxin Chen. Prior to that, he received his B.E. in Electronic Engineering from Tsinghua University in 2016. His research interests include high-dimensional data analysis, statistical machine learning, convex and nonconvex optimization, reinforcement learning, and information theory. He is particularly interested in developing scalable algorithms for data integration and information extraction, and understanding the theoretical foundations underlying the approaches, with an aim to achieve the optimal interplay between statistical accuracy and computational efficiency.
Hosted by: ECE Department
Submitted by: Olivia La Pierre <olivia@ece.ucsb.edu>