Postdoctoral researcher at the
Digital and Cognitive Musicology Lab
EPFL (Lausanne, Switzerland)
I am working in the area of machine learning and artificial intelligence with a focus on probabilistic modelling, planning, and active learning. While I am particularly fascinated by trying to understand human behavior, the involved questions are much broader, aiming at the comprehension of acting and learning systems in general. This knowledge is also relevant if we want to build artificially intelligent systems.
When we perceive the world we learn something about it. However, perceiving the world does not mean to passively take in all information. For example, already on the level of our retina the visual input is preprocessed and transformed (things like edge detection happen here to some extend) such that what actually reaches our brain contains more relevant information. These transformations change how the external world is represented internally. And what makes a good representation depends on what the information should be used for?
Perception is an active process in yet another way because our actions determine what we perceive and thus what we learn about the world. I can decide to open a box and learn what is inside or leave it closed and do something else instead. Learning and acting are thus intimately intertwined, which leads to the next question.
Taking the right decisions is difficult mainly for two reasons. First, we do not have perfect knowledge of the consequences of our actions, either because the world is inherently non-deterministic or because we are lacking some information about it. For choosing the best action we thus have to take into account all of its possible outcomes. Second, our actions change the world, which means that they potentially have effects in the remote future including what the effect of future actions is. It is therefore not enough to only consider the very next action but we effectively have to think about all possible sequences of actions and all of their possible outcomes – a task of overwhelming complexity. However, understanding the characteristics of our environment allows to come up with acceptable solutions even in situations that otherwise seem intractable, as the intelligent behavior of many humans and animals proves.
Finally, our environment is populated with other agents that we interact with. These agents, too, are able to perceive, learn, act, plan, and communicate. Perceiving ourselves as one agent among many others is an integral part of our human self-conception that we only just start to integrate on the technical level.
All of my research projects are, in some way or another, related to these three guiding questions.