Monday, November 21
12.10 - 1.00
CSE 4140

Speaker: Dilan Gorur, UC Irvine
Title: Scalable nonparametric Bayesian hierarchical representations

Abstract:
We are undergoing a data deluge. The amount of available data provides unprecedented opportunities to gain an understanding of the properties of the system that lead to data. Real phenomena is of essentially unbounded complexity and the increasing amount of data calls for models with increased capacity. The nonparametric Bayesian framework is a principled way to address this challenge by providing models of unbounded capacity that can flexibly adjust to available data. In this talk I will present a class of nonparametric Bayesian models for hierarchical representations, called Kingman's coalescent. The coalescent is a powerful model widely used in population genetics, and recent developments have begun to employ the approach as a hierarchical clustering model in machine learning. Although regarded as one of biology's most successful tools due to its generality and flexibility, the expense of coalescent-based inference has delayed the introduction of more model-based population genetic analysis. Similarly, lack of scalable inference techniques limited its use in machine learning to small scale problems. I will describe scalable greedy and sequential Monte Carlo algorithms for the coalescent that will make possible deployment of this model on a scale of data that was previously out of reach.

Speaker bio:
Dilan Gorur is a senior postdoctoral researcher in the Computer Science Department of UC Irvine. Prior to this, she was a postdoctoral researcher in the machine learning group at the Gatsby Computational Neuroscience Unit, UCL, London, UK. She did her PhD in machine learning in Max Planck Institute for Biological Cybernetics in Tuebingen, Germany in 2007. She received her MSc and BSc degrees from the Electrical and Electronics Engineering Department of METU, Ankara, Turkey. Her research interests are in theory and applications of machine learning, focusing on Bayesian nonparametrics.