Close this search box.

Zita Marinho Gives a Talk on Spectral Algorithms for learning Hidden Markov Models

Zita Marinho Gives a Talk on “Spectral Algorithms for learning Hidden Markov Models”

ZMarinho 2012 PRIBERAM Machine Learning Lunch Seminar
Speaker: Zita Marinho (CMU/IST)
Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação)
Date: Tuesday, March 26th, 2013
Time: 13:00
Lunch will be provided

In this seminar I will talk about the work of Hsu and Kakade (2009) on spectral methods for leaning a Hidden Markov Model (HMM). I will introduce and exemplify this method and describe how we can estimate the best set of state transitions or how we can predict a sequence of future observations. Hidden Markov Models are one of the most fundamental tools for modeling a discrete time series, and in general learning from an HMM is computationally hard, the classical approach to this problem resorts to search heuristics, like expectation maximization that are prone to local optima. I will talk about a new approach that learns an HMM based on the spectral decomposition of its parameters.This method depends implicitly on the number of distinct observations, making the algorithm particularly applicable to settings with a large number of observations, such as those in natural language processing.

Zita Marinho is a dual degree Ph.D. candidate in the CMU Portugal Program jointly advised by André Martins at Priberam/IST, Geoffrey Gordon at ML/CMU and Siddhartha Srinivasa at Robotics Institute/CMU. Her interests focus on learning methods from data observation and her PhD thesis is in spectral methods for learning in Natural Language and Robotics. She holds a Masters in science (CS) in Physics Engineering from IST, Portugal.

More information available at: