The four groups of the CI programme are all members of the Finnish Centre of Excellence in Computational Inference Research (COIN), and the objectives of the programme are closely intertwined with those of COIN.
The main objective of CI is to develop methods for transforming the data produced by the current data revolution into useful information. The key methodology for achieving this goal is statistical and computational inference based on the data. The emphasis is on large data collections and computationally demanding modelling and inference algorithms. Our mission is to push the boundary towards both more complex problems, requiring more structured data models, and towards extremely rapid inference. We plan to address a set of carefully chosen interdisciplinary “grand challenge” -level problems with high societal impact where solving the data intensive problems requires novel methodologies that can only result from combining the expertise of separate subfields.
Our mission brings up four partially overlapping methodological focus areas:
- Learning of massive data-driven models. This area, where our groups have a particularly strong track record, is about how to learn efficiently effective models from data. In the future we wish to further scale up our methods to “big data” problems involving massive data sets that may even require new, incremental or sub-polynomial methodological approaches.
- Learning from multiple sources. In the future the relevant information is typically distributed among a large set of partially connected sets. The parcelled and heterogeneous nature of the data is a problem, but is also an opportunity as while the data sets itself may contain mostly irrelevant data, the relationships between the data sources may give useful information, provided that suitable multi-source machine learning methods can be developed.
- Statistical inference in highly structured stochastic models. This focus area is concerned with the issue of how to do the modelling when the models are very complex in the sense of being structured by prior information and constraints. Despite the wealth of documented success stories about inference in highly structured stochastic models, awareness about the failure of standard stochastic computation techniques to handle increasing complexity of models is accumulating in the inference research community. In order to scale up existing stochastic learning methods, we will develop new types of intelligent adaptive algorithms.
- Extreme inference engine. This area is about real-time optimization and contraint reasoning problems: in particular in tasks related to contextual information retrieval, it is crucially important to get access to the information instantly and all the time, requiring on-line learning of the models and extremely rapid inference of what is relevant in the current context.
In applied research our work is also motivated by the big data and ubiquitous computing vision, where adaptivity, context-awareness and personalisation are key enablers. We see that our four methodological research areas support strongly each other, and they all address from a different perspective the key technological problems we face in our future “big data” information society. In our applied research we wish to link our strong basic research work in machine learning and constraint reasoning to well-motivated applied research activities involving prototype applications and real-world deployments.
- Bayesian Statistics, Professor Jukka Corander
- Complex Systems Computation, Professor Petri Myllymäki
- Computational Logic, Professor Ilkka Niemelä
- Statistical Machine Learning and Bioinformatics, Professor Samuel Kaski
Programme Director: Professor Jukka Corander
Programme Management Group
- Professor Petri Myllymäki
- Professor Samuel Kaski
- Professor Ilkka Niemelä
- Professor Jukka Corander