| This thesis introduces Hidden Process Models (HPMs). HPMs are a probabilistic time series model for data assumed to be generated by a set of processes, where each process is characterized by a unique spatial-temporal signature and a probability distribution over its timing relative to a set of known timing landmarks. Research on HPMs has been inspired and motivated by the functional Magnetic Resonance Imaging (fMRI) domain, and this document presents, develops, and evaluates this framework in the context of fMRI. We provide the HPM formalism, inference and learning algorithms, extensions to the basic formalism, a discussion of the correspondence between HPMs and Dynamic Bayesian Networks, experimental results evaluating HPMs on real and synthetic fMRI data, and examples of how to visualize the learned models. We conclude that the HPM extensions incorporating domain knowledge about the process signatures are important for analyzing real fMRI data, and suggest future improvements to the model. |
