EE/CS SEMINAR: Hierarchical Representations using Kingman's Coalescent

Faculty of Engineering and Natural Sciences
FENS SEMINAR 

"Hierarchical Representations using Kingman's Coalescent"

Dilan Görür, University of California Irvine 

Hierarchically structured data is found in a wide variety of domains such as vision (feature or object hierarchies), cognition (hierarchical categories, parts based learning), linguistics (parse

trees) and biology (phylogenies). Even when the data is not hierarchically structured, hierarchical representations are still useful simply as a statistical tool to summarize and visualize data and to efficiently pool information across the data at different scales. This talk concerns a class of hierarchical models constructed using Kingman's coalescent. We describe novel sequential Monte Carlo algorithms for inference on the class of models and present applications on learning visual taxonomies and hierarchical clustering of cell cycle using nucleus morphology.

Biography:

Dilan Gorur is a postdoc at University of California Irvine, and is also affiliated with California Institute of Technology. Prior to this she was a postdoc in machine learning at the Gatsby Unit for Computational Neuroscience, UCL. She did her PhD in Max Planck Institute for Biological Cybernetics in Tuebingen, Germany in 2003-2007. She received her MSc and BSc degrees from the Electrical and Electronics Engineering Department of METU in 2003 and 2000, respectively. Her current research is focused on theoretical and practical aspects of Bayesian machine learning applied to problems in various domains including neuroscience, bioinformatics, biology, cognitive science and computer vision.

 

12 January, 2010 at 13:40-15:00, FENS G032