Geometry of Diversity and Determinantal Point Processes: Representation, Inference and Learning
Speaker: Ben Taskar , University of Pennsylvania
Date: March 19 2012
Time: 4:00PM to 5:00PM
Location: 32-G449
Host: Regina Barzilay and Tommi Jaakkola, CSAIL
Contact: Francis Doughty, 253-4602, doughty@mit.edu
Graphical models are the dominant tool for capturing complex joint
distributions in computer vision and natural language processing
prediction tasks. However, inference and learning for all but a
small, restrictive subset of graphical models is intractable, and
standard approximation methods often fail for distributions with
global, negative correlations. Determinantal point processes (DPPs)
provide a computationally tractable alternative. DPPs arise in random
matrix theory and quantum physics as models of random variables with
negative correlations. Among their many remarkable properties, they
offer tractable algorithms for exact inference, including computing
marginals, computing certain conditional probabilities, and sampling.
DPPs are a natural model for subset selection problems where
diversity is preferred. For example, they can be used to select
diverse sets of sentences to form document summaries, or to return
relevant but varied text and image search results, or to detect
non-overlapping multiple object trajectories in video. I'll present
our recent work on a novel factorization and dual representation of
DPPs that enables efficient inference for exponentially-sized
structured sets. We develop a new inference algorithm based on Newton
identities for DPPs conditioned on subset size. We also derive
efficient parameter estimation for DPPs from several types of
observations. I'll show the advantages of the model on several
natural language and vision tasks: extractive document summarization,
diversifying image search results and multi-person articulated pose
estimation problems in images.
Joint work with Jennifer Gillenwater and Alex Kulesza, University of
Pennsylvania
Bio:
Ben Taskar received his bachelor's and doctoral degree in Computer
Science from Stanford University. After a postdoc at the University of
California at Berkeley, he joined the faculty at the University of
Pennsylvania Computer and Information Science Department in 2007,
where he currently co-directs PRiML: Penn Research in Machine
Learning. His research interests include machine learning, natural
language processing and computer vision. He has been awarded the
Sloan Research Fellowship, the NSF CAREER Award, and selected
for the Young Investigator Program by the Office of Naval Research
and the DARPA Computer Science Study Group. His work on
structured prediction has received best paper awards at NIPS and
EMNLP conferences.
See other events that are part of CS Special Seminar Series 2011/2012
See other events happening in March 2012
