LMSS @ Cornell Tech: Yonatan Belinkov (Technion)

Learning Machines Seminar Series What: LMSS: Yonatan Belinkov (Technion) When: Thursday, November 13, 1:30-2:45 pm Where: Bloomberg 081, Bloomberg Center, Cornell Tech (map) The series is organized by Associate Professor Yoav Artzi and sponsored by Bloomberg. Pizza will be served at 1:15 p.m. "Toward Scalable and Actionable Interpretability" Interpretability research has made many interesting discoveries about how language models and other deep learning models operate. However, despite this progress, interpretability has remained behind recent advances in how language models are used in practice. In this talk, I will describe some of our recent interpretability work, starting with scientific insights about the kind of algorithms that may be employed by language models. I will then describe several case studies where interpretability insights informed solutions for known problems: overcoming the modality gap in vision-language models and removing undesired information from trained models. If time permits, I will also share initial results on interpretability of protein language models. I will end by suggesting directions for making interpretability research more scalable and actionable. BIO Yonatan Belinkov is an Assistant Professor at the Technion. He is a former Azrieli Faculty Fellow and was a Mind Brain and Behavior Postdoctoral Fellow at Harvard University. Prior to that, he received his PhD from MIT. He is spending the current academic year at the Kempner Institute at Harvard University thinking about issues of interpretability, controllability, multi-agent communication, and AI for science.