Philipp Hummel

Philipp Hummel was a PhD student in the lab 09/21 – 05/23.

Despite its remarkable successes, deep learning still falls way short of a lot of human capabilities in terms of e.g. general-purpose problem solving, across-domain generalization, data efficiency, flexibly working with different levels of abstraction, etc. To make progress on these frontiers, I'm thinking about how to solve tasks involving higher cognition such as Bongard Problems (https://en.wikipedia.org/wiki/Bongard_problem) and tasks from the Abstraction and Reasoning Corpus (https://arxiv.org/pdf/1911.01547.pdf (wird in neuem Tab geöffnet)) with (probabilistic) program induction.

On another path I'm interested in uncovering the concepts and features that humans use more efficiently with the help of machine learning methods. This might be particularly useful for settings where we want to guide a learning model to use human-relevant instead of confounding features.

I am jointly supervised by Constantin Rothkopf and Frank Jäkel and funded by the KompAKI project. Before starting my PhD at TU Darmstadt, I studied Cognitive Science in Osnabrück and Neural Information Processing in Tübingen.