The MUPI group carries research on statistical machine learning and artificial intelligence. Our main goals are to
Statistical machine learning provides tools for understanding complex data collections, using Bayesian inference to cope with uncertainty on model parameters originating from learning from finite data. We develop computationally efficient and maximally automatic approximate algorithms for Bayesian inference, in context of probabilistic programming and machine learning. Our goal is to allow the user to focus on model specification, not needing to worry about the specifics of inference.
Hyperspectral cameras capture the full spectrum of light, instead of jus the three channels of red, green and blue that mimic the limited vision of humans. Having access to this richer information makes most computer vision problems easier. The existing HS cameras are, however, expensive and large. We develop a low-cost alternative that uses AI to process images captured with a passive add-on device that can be attached to any camera, brining HS imaging to smartphones and DSLRs.
For more details, see https://spexel.ai/.
Powerful targeted ultrasound can be used for non-invasive cleaning of, for example, industrial production equipment. We build machine learning and artificial intelligence tools for modeling ultrasound propagation in fouled structures, focusing in particular on methods that can be trained with very limited (or non-existent) direct observations of the fouling. The work is done in collaboration with the group of Edward Haeggström and Altum Technologies.
Many modern machine learning models are complex and require large training data sets, which makes learning difficult in applications where labeling examples is costly or difficult. We study data-efficient techniques for learning complex models from limited supervision, by utilizing related learning tasks (multi-task learning, transfer learning) and unlabeled observations or external constraints (semi-supervised learning). We also develop solutions for changing enrivonments based on domain adaptation.
Machine learning research is often carried out in the context of elegant but simplified setups: It is assumed that all relevant data is provided in form a simple matrix or tensor. In most practical applications this is not the case, but instead we need to combine information scattered in multiple data sources of heterogeneous nature. We provide fundamental modeling solutions for combining such data sources.
Today it is easy to collect information about individuals by monitoring their activities, either based on explicit sensors or by log data collected by computers they interact with. We develop models required for inferring interesting and useful information based on such data, to describe, understand and enchance our daily life.