Faculty Recruiting Support CICS

Learning from Unlabeled Video

06 Nov
Wednesday, 11/06/2019 11:45am to 1:00pm
Computer Science Building, Room 140
Machine Learning and Friends Lunch
Speaker: Carl Vondrick

I will discuss our research to use large amounts of unlabeled video in order to efficiently train models for visual recognition. Leveraging millions of videos, our work develops methods for machines to learn perception tasks such as anticipating human actions in the immediate future, tracking visual objects, and recognizing ambient sounds.  We show how to take advantage of the natural context available in video in order to learn without human supervision, for example through the natural synchronization of vision and sound, or the temporal coherence of motion and color.

Carl Vondrick is an Assistant Professor of Computer Science at Columbia University. Previously, he was a research scientist at Google. He completed his PhD at the Massachusetts Institute of Technology in 2017.