Author:

Abstract:

Sequential data (a.k.a. time-series) a rise in a multitude of different fields such as bi oinformatics, automatic speech recognition, roboti cs, computer vision, and computational finance. Wh ile the nature of the data generated in these appl ications is highly heterogeneous, the problems ari sing from such a wide range of fields reveal a str iking similarity when seen through the lens of pro babilistic time-series models. Most of the applica tion-specific problems fit in fact into one of the two main categories of algorithms developed for probabilistic time- series models: in ference (e.g. predict future values, estimate hidden variables from observab le ones, or partition a complex time-series in a s et of elementary segments) and learning (e.g. estimate time-series model para meters from observations, or discover pattern s or anomalies in time-series). In this thesis, a Bayesian treatment o f the uncertainties involved in time-ser ies models is adopted, and a set of probabili stic time-series models (that fall in the cat egory of switching models) is applied to four seemingly distant problems, namely¨ automatic segmentation of human gait time -series, classification of pathological g ait patterns, fault detection and recogni tion in robotic assembly tasks, and gas concentration estimation in unstructured envir onments using metal-oxide sensors. All these problems can be solved as inference or learning in a Bayesian time-series model, and have in common the crucial role that dom ain-specific expert knowledge plays in their¨ solution. The most attractive feat ure offered by the Bayesian framework is perh aps its potential to incorporate domain-s pecific expert knowledge either in the f orm of informative priors on the mod el parameters (to regularize the learning process) , or in terms of tailored conditional independ ence assumptions between model variables (to simplify inference). The main goal of this thesis is to investigate to whic h extent this potential can be fulfilled, and to h ighlight the scientific and pragmatic challenges i dentified and/or solved during this quest. In ¨particular, two Bayesian time-series models¨ are used in this thesis: the sticky-Hiera rchical Dirichlet Process Hidden Markov Model ¨(sticky-HDP-HMM) and the Augmented¨ Switching Linear Dynamical System model (aSLD S). The sticky-HDP-HMM is used in the gait segmentation and gait pattern cl assification applications. In both applications, r esults show that combining the HDP-HMM to an ad-hoc pre-processing step of the ga it time-series allows to transfer clinical kn owledge about human gait into the time-series model. This combined approach improves the p erformances in both applications, with respec t to the use of the HDP-HMM in isol ation. Motivated by these results, an extensi on of the HDP-HMM, named POLY-HDP-HMM is prop osed in this thesis. The sticky-HDP -HMM is also used for fault detection and recognit ion in robotic assembly tasks to learn force/torqu e sensor signature models. These mod els are used to classify faulty and succ essful task executions, allowing the robot to react on-line to errors and undertake er ror-specific recovery strategies. Finally, t he problem of estimating gas concentration in ¨unstructured envi- ronments using MOX sensor s is formulated as inference in an aSLDS model. In this application, domain-specific¨ expert knowledge is directly used in the construct ion of the model itself, rather than in estimating ¨its parameters. The proposed aSLDS model is¨ effective in overcoming the slow dynamical re sponse of MOX sensors, therefore extending their r ange of applicability in field robotics scenarios. In this thesis, we show that the adopted Bayesian time-series models can take advan tage of some of the available domain-specific expert knowledge to improve inferen ce and learning results. However, it appears¨ clear that, especially in clinical appli cations, further improvement can be achieved¨ increasing the transfer of clinical know ledge to the models. For this reason, we prop ose and discuss extensions to the time-series mode ls used in this thesis, and introduce new ones.