Event Location: NSH 1507
Bio: Ross Goroshin has recently obtained his PhD under Yann LeCun from the Department of Computer Science at New York University’s Courant Institute. He received a masters in Electrical & Computer Engineering from Georgia Tech and his bachelors in Electrical Engineering from Concordia University in Montreal, Canada.

Abstract: Many recent empirical successes in computer vision have been attributed to learning deep feature hierarchies by training on massive human-labeled datasets. Experiments have shown that despite being trained for specific tasks, feature hierarchies are transferable across applications. This observation naturally leads us to ask the following question: can generically useful feature hierarchies be learned without human supervision? This talk will explore the role of time as a weak form of supervision that can be used to train deep feature hierarchies with useful properties. The first part of the talk establishes a connection between Slow Feature Analysis and metric learning. It shows that temporally coherent features, trained on unlabeled videos, implicitly learn to be semantically discriminative. The second part of this talk presents a generative model that trains deep feature hierarchies to predict future video frames under uncertainty. It shows these features can implicitly linearize the underlying factors of temporal variation.