Research

The MUPI group carries research on statistical machine learning and artificial intelligence. Our main goals are to

  1. Develop computational techniques for learning under uncertainty, based on approximate Bayesian inference and probabilistic programming
  2. Solve interesting data-driven applications, focusing in problems short on data that hence need advanced modeling techniques based on data integration, semi-supervised learning etc.

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.

Highlights:

  1. Integration of approximate posterior inference and decision-making, by calibrating variational approximations to bettere account for the eventual decision task (Kuśmierczyk et al. Variational Bayesian Decision-making for Continuous Utilities, NeurIPS'19) and by post-hoc calibtration with flexible decision-making modules (Kuśmierczyk et al. Correcting Predictions for Approximate Bayesian Inference, AAAI 2020).
  2. Using prior predictive distribution for expert knowledge elicitation (Hartmann et al. Tackling uncertainty in prior predictive elicitation, UAI 2020) and for fast hyperparameter optimization of hierarchical Bayesian models (da Silva et al. Prior specification via prior predictive matching: Poisson matrix factorization and beyond, arXiv 2019).
  3. Flexible modeling of skewed data in probabilistic programming. (Klami et al. Lambert matrix factorization, ECML'18)
  4. Efficient model-independent variational approximation by re-using gradient computations during optimization. (Sakaya and Klami. Importance sampled stochastic optimization for variational inference, UAI 2017)
  5. Bayesian inference under combinatorial constraints. (Klami and Jitta. Probabilistic size-constrained microclustering, UAI 2016)

Projects:

  1. Scalable probabilistic analytics (SPA, 2016-2018). Development of computationally efficient variational inference algorihms for probabilistic programs. Funded by Tekes, Reaktor, Ekahau and M-Brain. Collaboration with Petri Myllymäki and Teemu Roos.
  2. Reliable Automatic Bayesian Machine Learning (RAB-ML, 2018-2019). Development of reliable and efficient solutions for Bayesian machine learning and probabilistic programming. Funded by Academy of Finland ICT-2023 program. Collaboration with Aki Vehtari and Antti Honkela.
  3. Finnish Center for Artificial Intelligence (FCAI), Agile probabilistic AI research theme.

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/.

Projects:

  1. Mobile hyperspectral imaging and computer vision platform (2019-2020). Development of methods and algorithms for acquisition of hyperspectral images with mobile devices, their use for computer vision, and preparation for commercialization of the result. Funded by Business Finland under the New Business from Research Ideas (TUTLI) instrument.

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.

Highlights:

  1. First example of AI model for detection and monitoring of fouling during ultrasonic sensing, based on monotonicity. (Rajani et al. Detecting industrial fouling by monotonicity during ultrasonic cleaning, MLSP'18).
  2. Detection and fouling quantification with Gaussian processes using only line integral observations. (Sillanpää et al. Ultrasonic Fouling Detector Powered by Machine Learning, IEEE International Ultrasonics Symposium (IUS), 2019; Longi et al. Sensor Placement for Spatial Gaussian Processes with Integral Observations, Uncertainty in Artificial Intelligence (UAI) 2020).
  3. Neural networks for localization internal structure with chaotic cavity. (Sillanpää et al. Chaotic cavity based acoustic location detector augmented with artificial intelligence, Physics Days, 2019).

Projects:

  1. Machine learning for ultrasonic cleaning (ML-UC, 2019-2023). Development of machine learning and artificial intelligence solutions for modeling ultrasound propagation in fouled structures. Funded by Academy of Finland.

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.

Highlights:

  1. Structured pseudo-labels for semi-supervised learning based on output-space constraints or smoothness assumptions. (Longi, Pulkkinen and Klami. Semi-supervised convolutional neural networks for identifying Wi-Fi interference sources, ACML'17).
  2. Transfer learning in computer science education. (Lagus et al. Transfer learning methods in programming course outcome prediction. ACM Transactions on Computer Education, 2018:18(4)).
  3. Segmentation of large multispectral satellite images based on coarse annotations in limited-memory computational architecture. (Luotamo et al. Multi-scale Cloud Detection in Remote Sensing Images using a Dual Convolutional Neural Network, arXiv preprint).

Projects:

  1. Scalable probabilistic analytics (SPA, 2016-2018). Development of computationally efficient variational inference algorihms for probabilistic programs. Funded by Tekes, Reaktor, Ekahau and M-Brain. Collaboration with Petri Myllymäki and Teemu Roos.

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.

Highlights:

  1. Bayesian canonical correlation analysis and inter-battery factor analysis (IBFA) and their extension Group factor analysis (GFA) for discovering relationships between more than two parallel data sources. (Klami, Virtanen and Kaski. Bayesian canonical correlation analysis, JMLR, 2013; Klami et al. Group factor analysis, IEEE Transactions on Neural Networks and Learning Systems, 2015).
  2. Matrix factorization tools for discovering relationships between more complex setups and heterogeneous data. (Klami, Bouchard and Tripathi. Group-sparse embeddings in collective matrix factorization, ICLR 2014; Klami. Polya-gamma augmentations for factor models, ACML 2015)
  3. Cross-domain object matching for discovering relationships between data sources with no known pairing between objects. (Jitta and Klami. Few-to-few cross-domain object matching, Advanced methodologies for Bayesian networks, 2017; Klami. Bayesian object matching, Machine Learning, 2013).

Projects:

  1. Traces of Information: Intelligence from Fragmented Sources (ToI, 2013-2019). Academy Research Fellow project of Arto Klami.
  2. Improved Learning by Combinin Information Sources (ILCIS, 2013-2015). Funded by Xerox Research Foundation, collaboration with Abhishek Tripathi (Xerox Research Center India) and Petri Myllymäki.

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.

Highlights:

  1. Modeling keyboard usage in programming education and other educational contexts. (Leinonen et al. Automatic inference of programming performance and experience from typing patterns, ACM TSCSE'16)
  2. Models for complex movement trajectories in indoor spaces. (Jitta and Klami. Partially hidden Markov models for privacy-preserving modeling of indoor trajectories, Neurocomputing 2017)
  3. Human computer interaction in touch-based information retrieval. (Andolina et al. Intentstreams: smart parallel search streams for branching exploratory search, IUI'15)
  4. Understanding intentions based on brain signal analysis, in particular MEG recordings during natural tasks. (Kauppi et al. Towards brain-activity-controlled information retrieval: Decoding image relevance from MEG signals, NeuroImage 2015)

Projects:

  1. Traces of Information: Intelligence from Fragmented Sources (ToI, 2013-2019). Academy Research Fellow project of Arto Klami.