BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//NextToU//EN
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:89315c9b-25c0-4c96-aeb7-1af9aff60092@nexttoyou
DTSTAMP:20261105T160747Z
DTSTART:20261105T160000Z
DTEND:20261105T170000Z
SUMMARY:Thesis Defense - Onkar Gujral
LOCATION:Building 4 4-146\, 182 MEMORIAL DR (REAR)\, Cambridge\, MA 02139
DESCRIPTION:Speaker: Onkar Gujral\n\nTitle: Mechanistic interpretability o
 f science-AI models\n\nAbstract:\n\nAI models trained for biological task
 s\, such as protein language models\, get used for a wide variety of high
 -stakes downstream applications like drug discovery. But we can’t see wha
 t these models are doing internally\, which makes them opaque blackboxes\
 , hindering trust and explainability in applications. We will describe ou
 r work on applying sparse dictionary learning methods to interpret protei
 n language models. This approach disentangles their internal representati
 ons 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 o
 n the mathematical task of distinguishing knot classes\, and then probe t
 he learned embedding space to reveal a surprising emergent mathematical r
 elationship. Finally we prove a result that extends the functoriality of 
 Khovanov homology to immersed surfaces.\n\nhttp://next-to-u.app/events/89
 315c9b-25c0-4c96-aeb7-1af9aff60092
URL:http://next-to-u.app/events/89315c9b-25c0-4c96-aeb7-1af9aff60092
END:VEVENT
END:VCALENDAR