AI Seminar sponsored by Fortive

Carl Vondrick -- Learning from Unlabeled Video

Abstract: Unlabeled video is available at massive scales and divulges the realistic complexity of everyday visual dynamics. In this talk, I will discuss our research to capitalize on unlabeled video to train computer vision systems without a human teacher. By creating learning algorithms that use incidental and structural clues naturally available in video, our research shows how to train computers to track objects, recognize human action, and anticipate future outcomes. Visualizations and experiments show that, although the models are not trained with ground-truth labels, rich perceptual representations emerge, which can be transferred across visual analysis tasks. We believe self-supervised learning is a promising approach to train machines to perceive their surroundings.

Bio: Carl Vondrick is an Assistant Professor of Computer Science at Columbia University. Previously, he was a Research Scientist at Google. He obtained his PhD from MIT. For more details, see his homepage.