Keynote Yonatan Belinkov, Technion - (Actionable) Interpretability beyond Language

Despite the increasing interest in interpretability over the past few years, the field has struggled to keep up with the frontier of AI. This talk will make two arguments. First, that for interpretability to be actionable, it needs to scale up. Second, that the community should pursue interpretability beyond natural-language models. The talk will include examples of (actionable) interpretability for unlearning undesired knowledge in language models and (not-yet-actionable) interpretability of protein models.