Document

Abstract

This Bachelor’s thesis discusses the rich theory of probabilistic graphical models and their relationship to structured data. Hidden Markov models are derived from the causal, directed and acyclic Bayesian networks. Conditional random fields are instead derived from the pseudo-probabilistic and undirected Markov random fields. The relationship between these models and more general structures are made clear. Inferential algorithms are discussed, tackling each of the fundamental HMM problems. Generalisation of these models to general topologies is achieved via the belief propagation algorithm. Parameter estimation and learning paradigms are discussed, with specific attention the Expectation-Maximisation principle; specifically, Baum-Welch learning. Finally, the use of PGMs to large vocabulary continuous speech recognition is discussed, specifically with regards to the TIMIT dataset. Limited experiments are preformed with a variety of graphical topologies. The results verify the validity of these models to automatic speech recognition, and indicate how PGMs might be used for other structured machine learning tasks.

Citation

1@thesis{verhoevenPGMsSequentialData2020,
2  type = {{{BSc}}},
3  title = {Probabilistic Graphical Models for Sequential Data},
4  author = {Verhoeven, Ivo},
5  year = {2020},
6  month = may,
7  address = {Middelburg},
8  school = {University College Roosevelt},
9}