[Colloquium] TTIC Talks: Borja Balle Pigem

[Dawn Ellis]

When:  Monday, April 8th at 11am

Where: TTIC-6045 S. Kenwood Ave.  Room 526

Speaker: Borja Balle (UPC)

Title: Spectral Methods for Learning Finite State Machines


Abstract: A simple algorithm for learning Hidden Markov Models via an
SVD decomposition was presented by Hsu, Kakade, and Zhang in COLT
2009. Following this seminal work, their spectral method has been
extended to many other probabilistic models, including: Predictive
State Representations, Kernelized Hidden Markov Models, Latent Tree
Graphical Models, High-dimensional Mixture Models, Finite State
Transducers, and Probabilistic Context-Free Grammars. The main feature
of the method is that it provides a local-minima-free alternative to
EM which is fast, easy to implement, and amenable to theoretical
analysis.

In this talk I will give a general and elegant description of this
spectral method in a framework that encompasses some of the results
mentioned above. In particular, those on models that can be defined in
terms of Finite States Machines. I will also discuss recent
applications and variations of this method that involve convex
optimization and matrix completion.

(based on joint work with X. Carreras, M. Mohri, and A. Quattoni)

-- 
*Dawn Ellis*
Administrative Assistant
773-834-1757
dellis at ttic.edu

TTIC
6045 S. Kenwood Ave.
Chicago, IL. 60637
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20130405/0f95b6d5/attachment.htm