About the event
This virtual session examines AI interpretability at two levels. The first talk shows how evaluating every step an AI agent takes, not only its final score, reveals safety and reliability risks. The next two look inside the models themselves: how a small set of visual tokens carries an image’s meaning into a vision-language model, and why a single task in a language model can have several equally valid explanations.
Disclaimer: This event is open to Vector Sponsors, FastLane companies, Vector Researchers, and invited partners only. Any registration that is found ineligible will be asked to provide verification and, if unable to do so, will not be able to attend the event. Please contact learn@vectorinstitute.ai with any questions.
Agenda
10:00 - 10:05 AM ET | Welcome
10:05 - 11:00 AM ET | Responsible AI Beyond Accuracy: Fairness, Safety and Sustainability (Talk + Live Demo)
Shaina Raza, Applied Machine Learning Scientist, Vector Institute
Aravind Narayanan, Associate Applied Machine Learning Specialist, Vector Institute
11:00 - 11:25 AM ET | To Sink or Not to Sink: Visual Information Pathways in Large Vision-Language Models
Jiayun (Letitia) Luo, PhD Candidate, University of British Columbia
Wan-Cyuan (Chris) Fan, PhD Candidate, University of British Columbia
11:25 - 11:50 AM ET | All Circuits Lead to Rome: The Intrinsic Multiplicity of Mechanistic Explanations in LLMs
Xi (Tony) Chen, PhD Candidate, University of Toronto
11:50 AM - 12:00 PM ET | Wrap-up