Public defence in Computational Science, M.Sc. Akash Kumar Dhaka
Opponent: Associate Professor Maurizio Filippone, EURECOM, Sophia Antipolis, France
Custos: Associate Professor Aki Vehtari, Aalto University School of Science, Department of Computer Science
The thesis is publicly displayed 10 days before the defence in the publication archive Aaltodoc of Aalto University.
Public defence announcement:
The task of finding distribution of unobserved variables given observed data is known as inference. Bayesian modelling offers a coherent and consistent approach of using probabilities to quantify uncertainty in inference, which in turn is propagated to output space in the form of a distribution over outcomes, in place of a point estimate. These models are being used in increasingly complex applications. The challenge however is that the procedure required requires solving high dimensional integrals, which is intractable in general. This means that there is no way to solve them exactly using pen and paper. This necessitates the application of approximate inference techniques. The quality of any analysis is determined by the success of the approximation inference algorithm chosen by a user.
Variational inference has emerged as a popular algorithm which due to recent innovations like automatic differentiation software and algorithmic improvements like stochastic optimization can be applied in a model agnostic way. The users are able to freely formulate probabilistic models without being constrained by limitation of inference algorithms. The first part of the thesis applied VI as inference technique to scale probabilistic Gaussian process models to large data or complex problems. The second part of the thesis shows that current optimization based algorithms struggle with posteriors having difficult geometry, provides some diagnostics and suggests some ways to improve upon current practices with some success.
Contact details of the doctoral student: [email protected], +358413132909