Agile probabilistic AI

(The FCAI research programs are currently in a ramp-up phase. More information will be updated here later.)

The goal of FCAI’s research program Agile probabilistic AI is to develop an interactive and AI-assisted process for building new AI models with practical probabilistic programming. The models will work as explainable, verifiable, uncertainty-aware, reliable tools to build and check the behaviour of AIs. Probabilistic programming has a vast set of applications in industry, e.g. in marketing predicting customer flows and in health care predicting medical treatment. Agile probabilistic AI contributes to mainly FCAI research objectives Data efficiency (objective I) and Understandability (objective III), and improves also trust (objective II) on the models and tools we have.

Our main expertise is in Bayesian machine learning, and we collaborate actively with other global leaders including DTU, Cambridge, and Columbia in NYC. We have recent core contributions in a wide range of inference algorithms, including expectation propagation, variational approximation, Gaussian graphical models, filtering and smoothing, approximate Bayesian computation, and methods for algorithm and model checking. FCAI also contributes to Stan, and other platforms, for 1) development of high-quality verified software for probabilistic programming, 2) development of computationally efficient and accurate model-agnostic Bayesian inference algorithms, and 3) theory and tools for understanding the modeling workflow as a whole, integrating model specification, validation, refinement, and visualization.

Coordinating professor: Aki Vehtari aki.vehtari at


The groups of following professors already take part in the research program Agile probabilistic AI. The list is currently under construction. If your group is already involved and needs listing here, please contact the program coordinator.