Spectral estimation of hidden Markov models

摘要: This thesis extends and improves methods for estimating key quantities of hidden Markov models through spectral method-of-moments estimation. Unlike traditional estimation like EM Gibbs sampling, the set methods, which we call HMMs (sHMMs), are incredibly fast, do not require multiple restarts, come with provable guarantees. Our first result upon original algorithm by parameters from fully reduced data. We also show that developed in dimensional version can be estimated using various forms regression, lead to major speed gains, as well allowing flexibility scheme. then extend beyond basic latent variable tree structures have linguistic applications, especially dependency parsing, finally output is a high-dimensional, continuously distributed variable. factored into two componentsestimation state space dynamics, observation probability distributions. leads extremely flexible procedures tailored precisely task interest. These tools all simple implement, naturally incorporate dimension reduction, allows them scale gracefully data increases. Degree Type Dissertation Name Doctor Philosophy (PhD) Graduate Group Statistics First Advisor Dean Foster Subject Categories Probability dissertation available at ScholarlyCommons: http://repository.upenn.edu/edissertations/1423 SPECTRAL ESTIMATION OF HIDDEN MARKOV MODELS