Title Of the Talk: The Missing Foundations of Interpretable Machine Learning
Speaker:Dr. Suraj Srinivas
Host Faculty: Dr. Konda Reddy Mopuri
Date & Time: Wednesday, 15th Nov 9:00am
Seminar link: https://meet.google.com/kop-vrti-hca
In this talk, I will present recent advances on the foundations of interpretable machine learning, highlighting fundamental conceptual bottlenecks. In particular, I will discuss the following issues with interpretability research: (1) the lack of a formal mathematical framework describing interpretability, obscuring its conceptual goals; (2) the lack of well-defined evaluation metrics, and in particular, a lack of a “ground truth”, making progress difficult to measure; and (3) the difficulty in distinguishing between plausible explanations (that aim to convince humans of model correctness) and faithful explanations (that aim to reflect model behaviour accurately). I will also present recent efforts at overcoming these issues, and highlight open problems, particularly relevant in the context of recent advances in large language models.
Suraj Srinivas is a postdoctoral research fellow at Harvard University, where he works with Prof. Hima Lakkaraju. He completed his PhD at Idiap Research Institute & EPFL, Switzerland, with Prof. Francois Fleuret, and his thesis was awarded the EPFL EDEE thesis distinction award. His research interests include designing algorithms and conceptual foundations for interpretable, robust, and computationally efficient deep learning.
Wednesday, 15th November 2023, 9:00 am