Eindhoven Area
Intelligence is more than just computation at scale - it's about making sense of the world under uncertainty. My work draws on probabilistic modelling and distributed inference to build machines that learn, adapt, and act like biological intelligent systems.
Dept. Electrical Engineering | Bayesian Intelligent Autonomous Systems lab I lead a research direction where we develop intelligent autonomous systems ("agents") for mobile robots. Our agents are designed based on the Free Energy Principle, a leading theory on neural information processing, and are implemented as message passing on probabilistic graphical models. Two key projects: FEPQuad, where we design agents for making quadrupedal robots learn to walk, and BayesBrain, where we develop a hybrid in-silico/brain-on-chip computer.
Department of Electrical Engineering | Bayesian Intelligent Autonomous Systems lab. I designed and developed Bayesian filters and smoothers for state, parameter and noise estimation in dynamical systems.
I developed proof-of-concept solutions based on research at TU/e and advised the Mathware department on Bayesian inference and probabilistic programming.
Datalogisk Institut | Image Analysis, Computational Modelling & Geometry section. I studied drifting data-generating processes, collaborating with the CopeNLU lab on tracking misinformation on social media and with the Medical Imaging lab on transferring knowledge of tissue labelling across medical centers.
I worked as a research software engineer on recognizing skull tissue in CT-scans for 3D-printed surgical assistive equipment (VUmc) and recognizing events in historical accounts of World War II (Meertens Institute / KNAW).
Department of Intelligent Systems | Pattern Recognition lab. Machine learning is about generalizing from observations, but future samples might differ from training samples due to biased sampling. My doctoral research focused on learning from one biased sample (source domain) and generalizing to another, differently biased sample (target domain). Firstly, I showed that standard model validation procedures often fail in domain adaptation settings. Secondly, I proposed statistical estimators that could adapt to domain shifts. Lastly, I formulated a “safe” estimator, i.e., one that is guaranteed to always perform at least as well as the baseline non-adaptive estimator.