Julian Christopher Kerl

I have studied Computer Science with a focus on machine learning and robotics at Technische Universität Darmstadt (Computer Science B.Sc. + Autonomous Systems M.Sc.). This included courses on the mathematical and technological foundations of computing, computer vision, machine learning, robotics, engineering, and many more. This included, for example, my Master's thesis, in which I developed an experiment in which I used computer vision and machine learning to make inferences about the behaviour of an artificial agent from the facial expressions of the people it was cooperating with. Additionally, I studied cognitive science (B.Sc. + M.Sc.), which focuses information processing in the (human) brain and included courses on psychology, perception ( especially vision), and neuroscience, as well as modelling, and machine learning. Part of this was, for example, a project where we used machine learning and computer vision to create a computational model of how humans perceive and process faces. During my time at TU Darmstadt I also worked as a student assistant at various departments. There I worked on a diverse range topics spanning from detecting and measuring cracks in steel and other materials using computer vision, to analysing political online communication and memes using multi-modal models and machine learning.

"Modelling curiosity constrained by planning" or "Solving problems before you know you have them"

In my PhD project, I want to look into the mechanisms that guide curious driven information collection in the context of planning problems. I am especially interested in the human ability to collect information that is not immediately relevant, but where collecting it now allows for solving planning more efficiently. My goal is to to gain insights into the mechanisms of information sampling and generalisation that underlay this remarkable behaviour, create models that allow us to better understand, and maybe even replicate it in machine learning agents.

In my work, I am interested in how humans curiously sample information without concrete knowledge of the problems they might need to solve with it. I want to gain insights into the mechanisms of curiosity that allow for this extremely efficient sampling, the mechanism of generalisation from little information that accompany it, and (ideally) develop models that can replicate this behaviour in machine learning agents. Outside of my main work, I am also interested in computer vision, human emotion recognition, and human behaviour.

Humans continuously collect information. A quick glance at something we haven't noticed yet on our way to work, taking a detour home, because it leads us through some part of the city we haven't explored yet, or even just skimming Wikipedia articles only tangentially related (if at all) to what we were actually trying to find out. The information itself often has no immediate use and is purely collected out of curiosity. Frequently, however, it ends up being very useful for solving novel or unanticipated problems. The fact that it was collected while there was seemingly no use for it, often proves especially beneficial. It not only means that the information is immediately available when it is needed, collecting it early also frequently ends up significantly reducing the cost of exploration (in terms of planning time, energy, resources etc.), compared to if we only started looking for that information when once we know what the problem looked like.