CS Guest Presentation: Trustworthiness In the Age of Foundation Models, Dr. Walter Gerych, Massachusetts Institute of Technology

Abstract:

Foundation Models (FMs) are large models that have been pretrained on an enormous corpus of diverse data, which are then adapted into a wide variety of downstream tasks such as object recognition and retrieval. As such, a single FM can affect many downstream systems, and thus flaws in the FM can be propagated to a range of systems. As they are increasingly deployed in high-stakes domains like healthcare and finance, ensuring the trustworthiness of FMs is crucial. Specifically, it is important to ensure that A) the FM is not relying on harmful biases for its decision making, and B) we are able to determine how confident we should be in any given output from the model. In this talk, I will discuss: 1) Minimally invasive methods for removing spurious correlations from FMs without harming their downstream performance. Any debiasing approach that degrades downstream performance is unlikely to be adopted in practice, necessitating debiasing methods that make as little change to the underlying model as possible. 2) Methods for obtaining robust measures of FM confidence. Specifically, I will discuss how to obtain measures for how confident an LLM is in its output, as well as how confident a user should be in the model itself.

Bio:

Walter Gerych joined MIT as a postdoctoral associate in 2023 and is hosted by Marzyeh Ghassemi in the Healthy ML group. He also serves as a Social and Ethical Responsibilities of Computing (SERC) Group leader at MIT, where he leads two teams of undergraduate and graduate students on projects focused on studying bias in clinical AI systems. Before coming to MIT, Walter completed his PhD in Data Science at Worcester Polytechnic Institute, where his dissertation focused on mitigating the negative effects of poorly labeled data and biased models. His current research continues to focus on making AI models less biased and more robust. He has a particular interest in debiasing large, pretrained models in practical and non-destructive ways. Beyond debiasing, he is also interested in promoting model robustness by developing calibrated measures of uncertainty in large language models and other AI systems. His primary application areas revolve around healthcare and mobile sensor time series for human activity recognition. He has also served as an organizer for the "Time Series for Health Workshop @ ICLR 2024" and as a track chair for "The AHLI Conference on Health, Inference, and Learning (CHIL) 2025". 

Faculty, undergraduate, and graduate students interested in AI are encouraged to attend!