Defence of doctoral thesis in the field of Computer Science, Eero Siivola

Statistical modelling is useful only if enough data is available about the phenomenon that is being modelled. The amount of needed data depends on the problem, but there usually is a cost in collecting it. This cost is not always measured in money as it can also be defined as time or wasted natural resources. The cost is particularly high in problems related to medicine, where the data might be collected from patients. Regardless of the modelling problem, it is desirable that as little data as possible is needed.

The dissertation further develops two methods for making statistical modelling more data parsimonious. The first method is to use new kinds of prior assumptions that are formed as data. The dissertation further develops methods that are based on assumed data, which is non-observed data, that is used similarly to the observed data to encode prior assumptions to the model. The second method that is presented in the dissertation is the intelligent gathering of data. This is usually done by iteratively selecting the next location where the data is gathered so that in light of the already gathered data it benefits from solving the modelling problem the most. The dissertation increases the understanding of these methods and extends them with new kinds of data that are based on ranking data at discrete locations.

The dissertation studies, compares, and further develops many already existing methods and develops new, computationally efficient approaches for solving the aforementioned problems using Gaussian processes as the statistical model. The proposed new approaches are general and are applicable to many different disciplines.

Opponent: Professor Sunil Gupta, Deakin University, Australia 

Custos: Professor Aki Vehtari, Aalto University School of Science, Department of Computer Science

Contact details of the doctoral student: [email protected], 0443933935

The public defence will be organised via Zoom. Link to the event

The dissertation is publicly displayed 10 days before the defence in the publication archive Aaltodoc of Aalto University. 

Electronic thesis