IC Colloquium : Efficient inference with combinatorial structure
Event details
| Date | 11.04.2013 |
| Hour | 15:00 › 16:30 |
| Speaker |
Stefanie Jegelka, University of California Berkeley IC faculty candidate |
| Location | |
| Category | Conferences - Seminars |
Abstract
Learning from complex data such as images or biological measurements invariably relies on capturing latent structure. But the combinatorial structure inherent in real-world data poses significant computational challenges for modeling, learning and inference, bringing commonly used probabilistic graphical models to their limits. In this talk, I will describe work towards overcoming these challenges. As a concrete example, I will introduce a novel, rich class of models whose inference procedures are not only efficient but also enjoy provable approximation bounds. This class of models combines two versatile concepts: graphs and submodular functions. While in its general form, this class can lead to hard inference problems, several practically relevant properties admit improved approximations or even fast exact inference. These results enable the realization of new models, such as a combinatorial sparsity that significantly improves image segmentation results in settings where state-of-the art methods fail. I will use this illustrative example to outline theoretical and empirical results, and will place them in a broader context within and beyond machine learning, sketching relations to structured sparsity, hierarchical models and implications for recently studied submodular optimization problems.
Biography
Stefanie Jegelka is a postdoctoral researcher at UC Berkeley, supervised by Michael Jordan and Trevor Darrell. She received a Ph.D. in Computer Science from ETH Zurich in 2012, in collaboration with the Max Planck Institute for Intelligent Systems and under the supervision of Jeff Bilmes and Bernhard Schölkopf. She completed her studies for a Diplom in Bioinformatics with distinction at the University of Tuebingen and the University of Texas at Austin. During her Diplom studies, she was a fellow of the German National Academic Foundation (Studienstiftung) and a member of its scientific college for life sciences. She also received a Google Anita Borg Europe Fellowship and has been a research visitor at Georgetown University Medical Center and at Microsoft Research. Her research interests lie in algorithms and machine learning.
Learning from complex data such as images or biological measurements invariably relies on capturing latent structure. But the combinatorial structure inherent in real-world data poses significant computational challenges for modeling, learning and inference, bringing commonly used probabilistic graphical models to their limits. In this talk, I will describe work towards overcoming these challenges. As a concrete example, I will introduce a novel, rich class of models whose inference procedures are not only efficient but also enjoy provable approximation bounds. This class of models combines two versatile concepts: graphs and submodular functions. While in its general form, this class can lead to hard inference problems, several practically relevant properties admit improved approximations or even fast exact inference. These results enable the realization of new models, such as a combinatorial sparsity that significantly improves image segmentation results in settings where state-of-the art methods fail. I will use this illustrative example to outline theoretical and empirical results, and will place them in a broader context within and beyond machine learning, sketching relations to structured sparsity, hierarchical models and implications for recently studied submodular optimization problems.
Biography
Stefanie Jegelka is a postdoctoral researcher at UC Berkeley, supervised by Michael Jordan and Trevor Darrell. She received a Ph.D. in Computer Science from ETH Zurich in 2012, in collaboration with the Max Planck Institute for Intelligent Systems and under the supervision of Jeff Bilmes and Bernhard Schölkopf. She completed her studies for a Diplom in Bioinformatics with distinction at the University of Tuebingen and the University of Texas at Austin. During her Diplom studies, she was a fellow of the German National Academic Foundation (Studienstiftung) and a member of its scientific college for life sciences. She also received a Google Anita Borg Europe Fellowship and has been a research visitor at Georgetown University Medical Center and at Microsoft Research. Her research interests lie in algorithms and machine learning.
Links
Practical information
- Informed public
- Free
- This event is internal
Contact
- Christine Moscioni