LMSS @ Cornell Tech: David Bau (Northeastern University)
- When
- Monday, November 9, 202612:00 PM – 1:00 PM ESTAdd to calendarGoogle Calendar
- Where
- Cornell Tech Bloomberg 0912 West Loop Road, New York, NY 10044Open in Maps
Reminders arrive as notifications from the NextToU iPhone app. Get the app
About
Learning Machines Seminar Series
What: LMSS: David Bau (Northeastern University) When: Monday, November 9, 12:00 p.m. - 1:00 p.m. Where: Bloomberg 091
The series is organized by Associate Professor Yoav Artzi and sponsored by Bloomberg.
Pizza will be served at 11:45 a.m.
"Interpretability, Indirection and Introspection"
When an LLM tells you about its own thoughts, does it really know them? The words "I know" or "I think" or "I want" are supposed to describe your own internal thoughts, but is it even meaningful to impart any meaning to such introspective language in an LLM? Does it even have any internal thoughts to describe? In this talk I will discuss some of the latest interpretability research in my lab, providing a brief overview of our work on in-context learning and binding. Then I will delve into our newest work studying the phenomenon and neural mechanisms underlying LLM introspection. I will describe a very simple experimental procedure for imparting new knowledge into an LLM that allows us to measure whether it can introspect. Then I will show how you can use this method to conduct a series of interpretability experiments to localize and relate introspection to other task performance. Our main question is: can LLM introspection be grounded, looking inward at causl physical neural states in the model? I will share insights about the relationship between introspection and model size, metalearning, and training, and I will discuss the implications of our work for AI safety and auditing.
BIO
David Bau is Assistant Professor at Northeastern University and Director of the National Deep Inference Fabric. His research studies the interpretability of emergent internal mechanisms in large neural networks accross both Natural Language Processing and Computer Vision. He received his PhD from MIT and is a recipient of the Sloan Research Fellowship.
Creator of ROME, Network Dissection, and GAN Paint. Published in PNAS, Neurips, ICLR, CVPR, SIGGRAPH, ECCV, EMNLP, ICCV. Formerly at Google. Co-author of the widely used textbook Numerical Linear Algebra.