PhD position on Learning Concepts with Theoretical Guarantees Using Causality and RL at University of Amsterdam

The Amsterdam Machine Learning Lab (AMLab) group at the university of Amsterdam is looking for a PhD candidate on this topic. The position is part of the Hybrid Intelligence consortium, a network of excellence of universities and institutes in the Netherlands focused on the combination of human and machine intelligence. This project will be in collaboration with the Sequential Decision Making group at TU Delft.

An AI model can have perfect accuracy on the training dataset even by learning the wrong concepts, i.e. by exploiting spurious correlations or other learning shortcuts. For example, a classifier might have perfect accuracy for the class “Cow” on a dataset in which we have only pictures of cows in Switzerland, but might fail miserably when applied to cows in the Netherlands, since it had used the background of mountains to classify the “Cow” concept. Learning transferable and robust concepts that can be reused or even composed in new settings requires reducing or ideally eliminating these shortcuts.

This PhD position will focus on addressing these issues by providing theoretical guarantees on learning interpretable high-level concepts from unstructured data, e.g. visual inputs, and actions in embodied AI settings. In particular, we will assume that high-level concepts correspond to ground truth causal variables. We will then develop a principled framework that combines causal representation learning and reinforcement learning (RL) to identify them by performing actions in an interactive environment with theoretical guarantees in terms of error and generalization bounds. Finally, we will apply the methods developed for this project in a real-world case study within the Hybrid Intelligence consortium on a Robotic Surgeon, an automated exoscope on a robotic arm that collaborates with surgeons.