As models become more complex, performing Bayesian inference is getting harder.
To tackle these new challenges, we currently explore two directions:

• Variational inference using normalizing flows

• Hybrid approaches that combine maximum-a-posterior solutions with MCMC sampling for Bayesian Neural Networks.

Slides of Invited Talk at BIRS,   Slides of Tutorial at BIRS,  Related Works

Accurate outlier detection is not only a necessary preprocessing step, but can itself give important insights into the data. However, especially, for non-linear regression the detection of outliers is non-trivial, and actually ambiguous. 

Here we develop a new methodology for handling outlier for Gaussian Process Regression that is complementary to the standard solution of using a Student-t likelihood.

Related Works

Deep learning, and in particular, large language models (LLMs) have become the dominant approach for text classification. However, explaining why the deep learning model decided a certain class is hardly possible. This is especially problematic for high-stake decisions in medical care.

As a solution, we develop new methods that can classify and analyze large amount of texts while providing simple explanations to the user.