Modeling, Prediction, and Diagnostics for Trustworthy AI
Rapid developments in communications, networking, robotics, genomics, new materials, and powerful computation platforms are bringing data-generating people, processes and devices together. The integration of data analytics in many fields are exciting because they enable new, faster and semi-automated methods in various areas of health, science and technology.
In this talk I will first briefly introduce some of my research in learning from online streaming data that may exhibit nonlinear, sparse, heterogeneous, time-varying, or recurring patterns. Challenges and recent advances in nonlinear modeling, variable selection, change detection, multi-regime analysis, rare event prediction will be introduced from a unified perspective. I will then focus on a particular recent advance in model selection. In particular, a new information criterion will be introduced to achieve an optimal learning performance when the candidate models may or may not be well-specified. It resolves a long-lasting challenge which the state-of-the-art model selection techniques including Akaike information criterion (AIC), Bayesian information criterion (BIC), and a variety of cross validations fail to address.
In the context of ‘machine learning fairness’, ‘statistical reproducibility’, and ‘trustworthy AI’, I will also address some misleading folklore concerning model comparison, parameter tuning, and cross validation in daily practice, guided by foundational theory. I will conclude the talk with some latest research on AI privacy and theory of neural networks.
Prof. Jie Ding
Assistant Professor, University of Minnesota on October 25, 2019 at 11:45 AM in EB2 1230
Jie Ding obtained his Bachelor's degree from Tsinghua University in 2012 and Ph.D. degree from Harvard University in 2017. He was a Postdoc Fellow at Harvard University in 2017 and a Postdoc Fellow at Duke University in 2018. He joined the faculty of the School of Statistics at UMN in August 2018.
His research interests are in foundational principles and efficient algorithms concerning statistical methodology, signal processing, and machine learning. His recent research focus on reproducible learning, prediction of online streaming data, and large-scale collaborative learning.
The Department of Electrical and Computer Engineering hosts a regularly scheduled seminar series with preeminent and leading reseachers in the US and the world, to help promote North Carolina as a center of innovation and knowledge and to ensure safeguarding its place of leading research.