Join us for Dr. Keith Burghardt's seminar hosted by UNC Charlotte's Center for Trustworthy Artificial Intelligence in Model Risk Management (TAIMing AI). 

Abstract:

Machine learning models make mistakes but solving the underlying reasons has been a 200-year challenge. I will review some common reasons why models fail guided in part by my industry experience and ongoing research. A core component I will emphasize is that common errors are not inevitable. Instead, current models make needlessly simplistic assumptions in order to be scalable. I will discuss three intuitive components of model bias: (1) features useful to some datapoints “leak” to other datapoints where they harm prediction; (2) models easily become distracted by spurious correlations; (3) models are not prepared for inevitable changes in test data. I then discuss potential solutions: (1) model architectures that “compartmentalize” feature leakage; (2) automated feature engineering to better capture relevant signals; (3) algorithms that automatically modify model parameters and test data to improve predictions. I end the talk by describing future work to better capture model errors.

Speaker Bio:

Keith Burghardt is an Assistant Professor at the University of North Carolina at Charlotte in the School of Data Science, where he researches scams and extremism in social media. Prior to joining the University, Keith was a research scientist at the University of Southern California's Information Sciences Institute as well as at Amazon.com. Keith is an editorial board member at Social Network Analysis and Mining and PLOS Complex Systems and has papers in several conferences, including WWW, ICASSP, ICWSM, ASONAM, AIES, and CSCW. His research has won awards such as Runner-Up Best Paper Award at ASONAM 2022, and has received extensive news coverage, including from Last Week Tonight with John Oliver, Nature Reviews Physics, The New York Times, Los Angeles Times, Wired, and CNBC. Burghardt received a PhD and BS (Magna Cum Laude with High Honors) in Physics at the University of Maryland in 2016 and 2012, respectively.

For more information, or to join our mailing list please email taiming-ai@charlotte.edu