Handling Uncertainty in Machine Learning Systems

The rapid development and deployment of machine learning systems is having a profound impact on our society. In recent years these systems have excelled in a wide range of data-driven prediction problems. However, in many cases, such as safety-critical applications, high predictive accuracy is not enough.

To make the systems robust we also need to reliably reason about the uncertainties in their predictions.

This project will result in new theory, evaluation methods, models and learning algorithms for handling uncertainty in machine learning systems. The project is organized into three tracks, each dealing with a certain aspect of uncertainty quantification.

The aim of Track 1 is to develop generic and interpretable tools for evaluating the reliability of predictive uncertainty.

Track 2 is about developing new learning algorithms and modeling approaches for ensemble-based reasoning about epistemic model uncertainty.

In Track 3 we will combine deep learning with probabilistic graphical models for encoding structure and dependencies in the model. Specifically, we will develop tools for robust uncertainty quantification in such combinations.

The tracks have been selected due to their high relevance for better understanding and handling uncertainty in machine learning, but also since they are connected.

A team of three junior researchers, supervised by the PI, will investigate the three tracks in parallel, thereby enabling positive synergies and cross-fertilization of ideas.