Department of Machine Learning

Mission

The key task of the Machine Learning Department is to carry out research into the theoretical rudiments of machine learning algorithms and related biologically inspired optimization algorithms, and to promote their use within the gamut of applications based primarily on data separation and on the prediction of time series in the field of science and engineering, and in other social spheres.

Our Research Area

As a subdivision of the computer sciences, machine learning deals with research into systems that learn solely on the basis of the knowledge of data. These systems currently form the backbone of artificial intelligence applications in diverse branches of science, industry, health care and sociology. An upsurge of artificial intelligence applications based on machine learning methods, a process evident in the past few years⁽¹⁾, has been facilitated by two aspects in particular. Firstly, by the wide-ranging availability of data describing the processes under scrutiny, which is also associated with the spread of digital scanning of information for its subsequent processing, and secondly by the significant progress in the performance of computer technology and its massive parallelization. These two factors have made it possible to use methods of artificial intelligence in a broad range of human activities, while the current growth of computing performance has caused that in some non-trivial applications artificial intelligence utilizing machine learning methods exceeds many times human capabilities (chess or the game go may be mentioned as graphic examples).

Many applications of machine learning are based on the use of computer structures described as deep neural networks, their variants and also on the optimization distributed methods of the genetic algorithm type. In spite of the indisputable practical success of this particular approach, the theoretical fundamentals of these methods have not yet been sufficiently explored. This tends to limit other possibilities for developing these applications to what is today the predominant empirical approach, which, however, lacks efficiency given by that exact approach based on profound knowledge of the principles on which the above-mentioned methods are known to be operating.

Researchers of the Machine Learning Department are focused on studying theoretical properties of machine learning methods, especially on exploring the properties of deep neural networks optimized by supervised learning, on examining different variants and designs of such models, and on clarifying related phenomena, among other things, convergence speed of learning, statistical reliability of learnt networks, robustness towards outliers or adversarial examples and other issues.

Vision

A long-term task facing the Machine Learning Department is to achieve in-depth and broad understanding of the machine learning principles and to use such knowledge for designing more efficient methods in applications based on machine learning.

In view of the existing indications that higher levels of human intelligence are activated by similar neural structures as cognitive abilities, it is anticipated that understanding of the actual principles of machine learning should, at the same time, be conducive to creating new theoretical branches of informatics that will capture those higher levels of human intelligence.

Selected papers

On convergence of kernel density estimates in particle filtering, 2016, D. Coufal

On locally most powerful sequential rank tests, 2017, J. Kalina

Implicitly Weighted Methods in Robust Image Analysis, 2012, J. Kalina

Human-Inspired Eigenmovement Concept Provides Coupling-Free Sensorimotor Control in Humanoid Robot, 2017, A. V. Alexandrov, V. Lippi, T. Mergner, A. A. Frolov, G. Hettich and D. Húsek

Comparison of Seven Methods for Boolean Factor Analysis and Their Evaluation by Information Gain, 2015, A. A. Frolov, D. Húsek, P. Y. Polyakov

Probabilistic lower bounds for approximation by shallow perceptron networks, 2017, V. Kůrková, M. Sanguineti

Constructive lower bounds on model complexity of shallow perceptron networks, 2017, V. Kůrková

Measures of ruleset quality for general rules extraction methods, 2015, M. Holeňa

Using Copulas in Data Mining Based on the Observational Calculus, 2015, M. Holeňa, L. Bajer, M. Ščavnický

Evolution Strategies for Deep Neural Network Models Design, 2017, M. Vidnerová, R. Neruda

Seminars

The department organizes a seminar HORA INFORMATICAE with both local and international speakers. Our colleague Martin Holeňa organizes Seminar of Machine learning and Modeling.

Topics for dissertation and diploma thesis

Estimates of the number of patterns in the PAC model for nonconsistent separation.

Analysis of data-mining methods based on univalent neural networks.

Robust classification for high-dimensional data.

Robust estimation in neural networks.

Projects

Metalearning for Extraction of Rules with Numerical Consequents, No. GA17-01251S
Research infrastructure for Fermilab experiments, No. LM2015068
Automated Knowledge and Plan Modeling for Autonomous Robots, No. P103-15-19877S
Model complexity of neural, radial, and kernel networks, No. P202-15-18108S

Awards and other

Bolzano Medal for Merit in the Mathematical Sciences, Czech Academy of Sciences, 2010 (V. Kůrková)
Best Paper Award, BIOSTEC/BIOINFORMATICS Conference, Rome, Italy, 2016, (J. Kalina, J. Hlinka)
Best Paper Award, ITAT, 2017, Slovakia, (P. Vidnerová, R. Neruda)
Famelab Olympics, gold medal, Dublin, 2012 (J. Kalina)
Editorial boards: Neural Networks, Elsevier (since 2009), Neural Processing Letters, Springer (since 1995), IEEE Transactions of Neural Networks (2008-2009), guest editor of the special issues of Neural Networks 23(4), 2010, and Neurocomputing 96, 2012.
Scientific committees: Board of ENNS (European Neural Network Society, since 2007), president of ENNS 2017-2019, Advisory Committee of SIREN (Societa Italiana Reti Neu- ronali, since 2009), Advisory Committee of ICANNGA (2001-2015), Management Committees of COST Actions IntelliCIS (2009-2013) and EuNetAir (2012-2016)
Invited plenary talks at conferences: ICANN 2016, Spain; WCIDM 2014, Slovakia; MENDEL 2013, CR; CINTI 2012, Hungary; ICAFS 2010, CR; IPDDM 2010, Austria; IP:MS 2010, Turkey; WIRN 2009, Italy
Tutorials at conferences: EANN 2015, Greece; EANN 2011, Greece; ICANNGA 2007, Poland
Patent No. 306533: Hardware implementation of neural nets with schwitching units (F. Hakl)

Media links related to AI and machine learning⁽¹⁾

Department Members

Head: František Hakl

Secretarial support: Iveta Kubíková