Department of Computer Engineering
S E M I N A R: Label Efficient Deep Learning for Robot Perception
Dr. Ozan Şener
Stanford University
Supervised learning with large scale labelled datasets and deep layered models has caused a paradigm shift in diverse areas in learning and recognition. However, we did not see a similar shift in robotics. Our robots are still having trouble recognizing basic objects, detecting humans, and even performing simple tasks. A popular conjecture for this limitation is the lack of large-scale data in robot perception. A plausable way to ease this problem is using simulators in addition to the real world data. However, this approach suffers from generalization issues under the presence of a domain shift between the training and the test data distribution (simulated data vs real data). Moreover, it is also crucial to have smart ways for choosing points to be labelled (i.e. active learning) in order to minimize labelling efforts in real data.
In this talk, we will first present a unified deep learning framework where the representation, cross domain transformation,and target label inference are all jointly optimized in an end-to-end fashion for unsupervised domain adaptation. We further develop a geometric reasoning explaining the success of our method by providing an upper bound for the generalization error. Using the insights obtained from this geometric understanding, we further propose an active learning algorithm carefully tailored for convolutional neural networks with theoretical guarantees. Our algorithms drastically reduces the annotation requirements by transferring knowledge from simulation and guiding the data collection process. We also apply our methods for various robot manipulation tasks.
Bio: Ozan Sener is a post-doctoral researcher at Stanford University. He earned his BS. and MS degrees from the Middle East Technical University. He joined Cornell University as a PhD student and later moved to Stanford University following his advisor’s move. His research interests include machine learning, computer vision and robotics. His recent works involve transfer learning among domains and modalities, unsupervised learning using generative models, and geometric approaches to deep learning.
DATE: 21 August 2017, Monday @ 13:40
PLACE: EA-409