Efficient Bayesian Inference and Computation
PIs: Luigi Acerbi, Arto Klami, Antti Honkela
We develop novel methods to make Bayesian inference more efficient, scalable, and practical. This includes work on variational methods, Monte Carlo algorithms, and techniques for handling complex models in both likelihood-based and simulator-based inference. Our approaches range from geometric methods to sample-efficient algorithms and meta-learning or amortized techniques, with applications spanning from cognitive science to privacy-preserving computation.
Trustworthy and Accessible AI
PIs: Antti Honkela, Kai Puolamäki, Teemu Roos
We develop methods to make AI systems more reliable, transparent, and accessible. This encompasses privacy-preserving machine learning, explainable AI, efficient learning algorithms, and techniques for uncertainty quantification. Our work ranges from fundamental algorithmic improvements to practical implementations, with applications in healthcare, education, and large-scale data analysis. We also focus on making machine learning more accessible through innovative teaching platforms.
Theoretical Foundations and Algorithms
PIs: Mikko Koivisto, Aapo Hyvärinen, Teemu Roos
We advance both theoretical foundations and fundamental algorithms in machine learning. This includes theoretical work on identifiability, causality, and computational complexity, as well as development of core algorithms for tasks like nearest neighbor search, clustering, and structured probabilistic models. Our research combines mathematical rigor with algorithmic innovation to enable efficient learning from data.
Neural Computing and Brain-Inspired AI
PIs: Aapo Hyvärinen, Luigi Acerbi
We study and develop neural network approaches with inspiration from and applications to neuroscience and cognitive science. This includes work on independent component analysis, self-supervised learning, the role of noise in neural networks, and probabilistic models of neural and behavioral data. Our research spans from fundamental principles of neural computation to practical applications in neuroimaging, brain signal analysis, and understanding human learning and decision-making through the lens of Bayesian inference.
Machine Learning for Scientific Discovery
PIs: Kai Puolamäki, Arto Klami, Luigi Acerbi, Aapo Hyvärinen
We develop machine learning methods to support scientific research across multiple domains. This includes novel approaches for exploratory data analysis, virtual laboratory frameworks for scientific experimentation, and computational modelling in the sciences. Our methods emphasize interpretability and uncertainty quantification to afford rigorous scientific inference. Applications include atmospheric and earth system sciences, industrial ultrasonic sensing, and cognitive neuroscience.