Convincing Experts to (not) Trust ML Models

Title: Convincing Experts to (not) Trust ML Models

Abstract: ML systems have a long history of being unreliable---should we trust these models today? On the one hand, ML systems are prone to manipulation, exemplified by jailbreaks for language models. On the other hand, ML systems may know more than scientists and doctors, as they can make inferences based on superhuman quantities of data. In this talk, we will discuss our work in building new models, theory, and benchmarks to enable expert trust in ML systems, with applications in safety alignment, cosmology, and surgery. 

Bio: Eric Wong is an Assistant Professor in the Department of Computer and Information Science at the University of Pennsylvania. He researches the foundations of reliable machine learning systems: understanding, debugging, and guaranteeing the behavior of data-driven models. In practice, his research empowers expert scientists and doctors to learn from AI models and make new discoveries. Eric received his Ph.D. in Machine Learning from Carnegie Mellon University, was a postdoctoral researcher at Massachusetts Institute of Technology, and is a recipient of the Siebel Scholarship, SCS Dissertation Award (honorable mention) and an Amazon Research Award, as well as paper awards at IJCNLP-AACL (area chair award) and NeurIPS workshop on ML & Security (best defense).