Rasmus Bonnevie: Exact inference in Bayesian models is often intractable, which has led practitioners to construct approximate Bayesian posteriors using MCMC or, more recently, variational inference.
Both strategies can fall short when faced with highly multimodal posteriors, as the approximations are updated locally and can fail to take global structure into account. In the case of variational inference, this is predominantly due to limitations on the expressivity of the approximation family imposed by classical inference methods, and I explore how these limitations can be circumvented.
More broadly, I also explore how inference can be efficiently extended to novel problems and how to make inference more robust for classes of flexible models such as Bayesian nonparametric mixtures and Gaussian processes.
Supervisors: Main supervisor Mikkel Nørgaard Schmidt, Co-supervisor Morten Mørup
Research section: Cognitive Systems
PhD project title: Structured Approximate Bayesian Inference
Published as PhD report: Structured Bayesian Approximate Inference