ADA draft non-public: challenges

(

Old:

• Example challenge 1: Learning Network Structures. Network-like structures are numerous in various domains including molecular processes, social interactions, and the Internet. New computational methods are needed for finding the structure of such networks and for understanding their dynamic behaviour.
• Example challenge 2: The Vocabulary, Grammar and History of Genomes. The genome codes information identifying the species and the individual. Computational techniques are needed for the description and the analysis of variation. Segmentation methods using recurrent sources can be used to find components with similar underlying structure; latent variable techniques for sequences can also be used.
• Example challenge 3: Computational Modelling of Ecosystems. The environment can be measured in many ways on different scales ranging from remote-sensing based satellite images of landscapes to chemical compositions of nutrients in individual plants. The complex interactions in both the spatial and temporal domains across different scales are largely unknown, and their importance is growing.
• Example challenge 4: Sensor and Context Data Management. To realize a vision of ubiquitous information processing, services and applications make use of a wide variety of context data, including sensor readings. The challenges are to efficiently gather sensor data, to perform context reasoning, and to take into consideration the resource constraints of the devices and the distributed nature of the environment.

New:

(Mannila) We develop new concepts, algorithms, principles and frameworks for algorithmic data analysis. We believe that developing new concepts and algorithms is at best an iterative process, consisting of interacting extensively with the application experts, formulating computational concepts, analyzing the properties of the concepts, designing algorithms and analyzing their performance, implementing and experimenting with the algorithms, and applying the results in practice. Our current application areas include biology, ecology and paleontology, visual analytics, and borders in various disciplines such as biology, ecology, history and geography.

(Toivonen) We develop novel methods and tools for pattern and link discovery. Our focus is on structured and heterogeneous data, such as graphs, and also on sequences. The importance of data mining in heterogeneous and structured data will only grow in the future. There will be an increasing amount of challenging and important problems, especially in scientific applications. Our main applications are in bioinformatics, in collaboration with applied scientists and companies. Our current emphasis is on analysis and link discovery in weighted (biological) graphs. We identify computational problems in them, develop new algorithms, and apply them. 

(Hyvärinen) Computational neuroscience is the construction of mathematical models of neural information processing. Because of their inherent complexity, neural circuits need theoretical models which abstract away unnecessary detail and allow us to investigate the crucial aspects of the computation. Also, our work has focused largely on the visual system.The approach is to consider how the brain performs a sophisticated statistical and probabilistic analysis of the environment. We analyze the statistical structure of natural images and image sequences. Statistical regularities in the data indicate what kind of features are optimal for processing this data. It happens that the optimal features are, in many cases, very similar to those that the visual cortex is known to compute. These methods also have applications in image processing and computer vision.

(Hollmen) The research group Parsimonious Modelling develops computational methods for data analysis and applies these methods on two particular application fields: cancer genomics and environmental informatics. Both of these application fields exhibit problems of high dimensional data and complex, unknown interactions between measurements. Parsimonious modeling aims at achieving maximally simple or compact models as a result of the data analysis process. In practical problems, parsimony makes results more understandable and interpretable. For instance, feature variable selection aims at parsimony in terms of the number of variables in the model. The computational methods are based on regularization or penalization of the cost function in the original process of learning from data and may be also combined with heuristic search procedures.

(Kaski) Combining the different kinds of current high-throughput data produces new systems-level hypotheses about gene function and regulation, and ultimately functioning of biological organisms. We develop probabilistic modeling, statistical data analysis and machine learning methods to advance this field.

(Kaski) Successful proactivity, i.e. anticipation, in varying contexts requires generalization from past experience. Generalization, on its part, requires suitable powerful (stochastic) models and a collection of data about relevant past history to learn the models. Our goal is to build statistical machine learning models that learn from the actions of people to model their intentions and actions. The models are used for disambiguating the users' vague commands and anticipating their actions. The actions and interests are monitored by measuring eye fixations and movements that exhibit both voluntary and involuntary signs of the users' intentions.

(Mäkinen) The research group studies a new subfield of data compression - data structure compression. The new aspect compared to traditional compression is that the compressed data (structure) needs to be represented so that access to its internal parts is provided without uncompressing the whole structure. As an example, consider a binary tree of n nodes. It is possible to represent the tree succinctly using about 2n bits so that the children and parent of any node can be accessed in constant time. A standard link structure representation of a binary tree takes of order n log n bits.