Bio

Towards AI-enabled Healthcare Through Learning Effective Representations Efficiently

Abstract

Learning effective representations efficiently plays a pivotal role in machine learning applications ranging from computer vision, and natural language processing, to healthcare, mobile sensing, and computational biology. Given a large amount of informative but also noisy data in various domains, one of the core quests for artificial intelligence is motivating the design and analysis of new representation learning methods for various applications. The research objectives of this talk are to understand and develop new representation learning models through the lens of optimization and measurement, apply representation learning frameworks for various applications, and invent new metrics to creatively evaluate representation learning outcomes. First, Adaptive Sampling with Reward and Deep Wasserstein Learning frameworks are presented to tackle the sampling and measurement challenges in ranking-based loss functions with applications in computer vision and single-cell biology. Second, I will present data-efficient learning methods for mobile sensing and cross-model single cell data through coreset. Next, I will present a case study using representation learning to understand lung cancer patients' treatment trajectories after immunotherapy. Lastly, three evaluation metrics based on the science of science, information theory, and causality respectively are presented for knowledge measurement in knowledge representations. Taken together, this talk outlines a highly comprehensive, impactful, and interdisciplinary approach to representation learning research, from the perspectives of models, applications, and metrics, with broader implications for artificial intelligence and machine learning more generally.

Bio:

Jason Xiaotian Dou is a postdoc research fellow at Harvard University working on artificial intelligence and medicine.  He did his Ph.D. in Computer Engineering at the University of Pittsburgh/University of Pittsburgh Medical Center (UPMC). His Ph.D. dissertation is titled “Learning Effective Representation Efficiently: Models, Applications, and Metrics.” Previously He did B.S. in Computer Science from Peking University with an undergraduate thesis at Carnegie Mellon University on “A Markov Chain Approach to the Gerrymandering Problem.” He also did master study in Information Science at Cornell University. His research is in machine learning, specifically representation learning with various applications in computer vision, natural language processing, single-cell biology, mobile sensing, and healthcare. His works have appeared in AI and machine learning conferences such as AAAI, Neurips, ICML, ML4H, and ICIBM and journals like Clinical Cancer Research. His work on Internet Finance Index has been featured in news media such as Tencent, Sina, and Caixin. For more about his works please go to his website jasondou.org.