Events
Colloquium
"Heuristic Estimation of Neural Network Outputs"
 |
Speaker: Mike Winer (IAS) Marvin L. Goldberger Member, School of Natural Sciences, Institute for Advanced Study Institute for Advanced Study Wednesday, October 22, 2025 11:30AM - 1:00PM Lunch will be served at 11:30am in room 1307
Talk will be 12:00-1:00pm in 1327 Location: Yale Institute for Foundations of Data Science, Kline Tower 13th Floor, Room 1327, New Haven, CT 06511 and via Webcast: https://yale.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=fa8695d3-331c-40fa-9720-b35201057c71 |
Talk summary: Given a neural network and a description of its input distribution, what can we say about the outputs? In some sense we have all the information, but even estimating something like the frequency of a given rare token might require many forward passes. In this talk I discuss approximate techniques for answering these questions, often in manners much more computationally efficient than blindly producing forward passes. I discuss how these techniques shed light not only on what neural networks do on a given input, but why they do it.
Speaker bio: Michael Winer is a statistical physicist who studies disordered systems, their phase transitions, thermodynamics, and dynamics. Much of his work focuses on the physics of glasses and how it connects to broader questions in holography, deep learning, and the emergence of complex behavior from simple components. He is interested in how systems of many simple parts can organize into phenomena such as magnets, glasses, or intelligence. Michael currently divides his time between the Institute for Advanced Study in Princeton and the Alignment Research Center in Berkeley.
Add To: Google Calendar | Outlook | iCal File
- Colloquium
Submit an Event
Interested in creating your own event, or have an event to share? Please fill the form if you’d like to send us an event you’d like to have added to the calendar.