Thesis Defense - Onkar Gujral
bis
Building 4, 4-146, Suffolk
Speaker: Onkar Gujral Title: Mechanistic interpretability of science-AI models Abstract: AI models trained for biological tasks, such as protein language models, get used for a wide variety of high-stakes downstream applications like drug discovery. But we can’t see what these models are doing internally, which makes them opaque blackboxes, hindering trust and explainability in applications. We will describe our work on applying sparse dictionary learning methods to interpret protein language models. This approach disentangles their internal representations into human-understandable biological concepts, improving downstream explainability and opens the door to learning new science trapped inside such models. In a similar vein, we also train a contrastive loss model on the mathematical task of distinguishing knot classes, and then probe the learned embedding space to reveal a surprising emergent mathematical relationship. Finally we prove a result that extends the functoriality of Khovanov homology to immersed surfaces.