Title: AI Interpretability: From Insight to Action 

Speaker: Yonatan Belinkov, Kempner Visiting Scholar and Assistant Professor of Computer Science at Technion - Israel Institute of Technology

Abstract: Interpretability research has made various discoveries about how language models operate. However, despite this progress, interpretability has remained behind recent advances in how language models are used in practice, and had little impact on shortcomings of these models. In this talk, I will describe some of our recent efforts to harness interpretability insights for actionable solutions. I will start with insights about the kind of algorithms that language models employ, highlighting how they use a bag of heuristics rather than implement robust algorithms. I will then describe several case studies where interpretability insights informed solutions for known problems, focusing on removing undesired knowledge from language models. Finally, I will discuss how acquired heuristics may help in scientific discovery, with protein language models as a test case. The talk will conclude with a call for action for the community to work on scalable and actionable interpretability.

Speaker Bio: Yonatan Belinkov is currently a Visiting Scholar at the Kempner Institute. He is an Associate Professor at the Faculty of Computer Science at the Technion – Israel Institute of Technology. Prior to that, he was a postdoctoral fellow at Harvard SEAS and a postdoctoral associate at MIT CSAIL. He received his Ph.D. in Electrical Engineering and Computer Science from MIT in 2018.