Programme of colloquia with abstracts for the Autumn 2011 semester

27 September 2011

Mgr. Vojtěch Malínek, Institute of Czech Literature, Czech Academy of Sciences, Prague

The Retrobi System

Abstract: The Retrobi system is being developed to present data from the card catalogue Retrospective Bibliography of Czech Literary Studies, which, with approximately 1.75 million cards, ranks among the largest in the Czech Republic. It handles the processing of input data from scanning right through to publication in a web application, i.e. in particular the automatic detection of blank pages, the merging of multi-card entries, and the enrichment of scanned images with their OCR transcriptions, as well as the associated migration processes and control mechanisms. In addition to the basic presentation of the catalogue’s image data in the form of simple browsing, the web application is enhanced with the option of full-text searching of OCR transcriptions, online correction of published text data by registered users, including an administrator interface for checking this data, a bulk editing function, and the ability to create and further process any user-defined searches within the system.

4 October 2011

RNDr. Tomáš Brázdil, Ph.D., Faculty of Informatics, Masaryk University

How to make optimal decisions in a stochastic environment

Abstract: In practice, we are often forced to make repeated decisions, the consequences of which we can only estimate approximately. Examples can be found in many fields, ranging from management through industrial process control to biological experiments. It appears that the quality of decisions can be significantly improved through the use of mathematical modelling and analysis. Markov decision processes provide a fundamental formalism for modelling precisely such decisions, which exhibit elements of quantified uncertainty; that is, we are only able to estimate the probabilities of possible outcomes. A simple example is betting on roulette, where, although we do not know the final profit in advance, we are able to estimate the probabilities of possible outcomes. An optimal decision can only be made if the values or prices of our actions (decisions) are defined. In this lecture, I will focus on decision processes in which the aim is to maximise a certain value of actions in the long term. The lecture will include a presentation of the latest results from the theory of Markov decision processes, whose actions are valued by multidimensional vectors of real numbers, which allows for optimisation from multiple perspectives.

11 October 2011

Assoc. Prof. RNDr. Tomáš Pitner, Ph.D., Faculty of Informatics, Masaryk University

Imagine Cup: A talent competition and the FI team’s successes at the global forum

Abstract: The Microsoft Imagine Cup (www.imaginecup.com) is the world’s premier student technology competition. It provides an opportunity for students to use their creativity, passion, and knowledge of technology to help solve global challenges. It is held annually in several categories, mostly for teams, but it also offers opportunities for talented individuals. The Celebrio Software team from the Lasaris research lab, Faculty of Informatics, succeeded in this year’s world finals in New York, finishing between 7th and 18th place in the highly competitive global field of 67 teams in the finals of the prestigious Software Design category. The colloquium will aim to convey some of the competition’s objectives, profile and spirit to the Faculty in order to inspire others to follow in Celebrio’s footsteps. In particular, the ‘real-life experiences’ relating to the evaluation process and criteria, as well as examples of other successful projects and trends, will be presented.

(The presentation will be delivered jointly with members of the Celebrio team, in Czech or English.)

18 October 2011

Prof. RNDr. Jiří Zlatuška, CSc., Faculty of Informatics, Masaryk University

Research evaluation in computer science – approaches and challenges

Abstract: Compared with those in older and more established scientific disciplines, possess certain distinctive features, manifested in the special role of conference papers in publishing new findings, as well as specific disciplinary characteristics of research that is not a purely theoretical discipline. The Czech methodology for evaluating research is unique, and its negative effects on the scientific system as a whole are documented not only by domestic critics but also, currently, by the findings of an international audit of research and development evaluation in the Czech Republic, which specifically highlights the negative effects of equating evaluation with decision-making (or, rather, replacing it) in the allocation of funding. The differences and methodological foundations of evaluation in computer science, which originate from the US National Research Council, the Computing Research Association and the European Informatics Europe, provide a set of recommendations that can be used both to identify inappropriate features of an evaluation system and to specify the requirements that a workable system should meet. Larger documented evaluations from abroad can serve as case studies of current practice in this field.

25 October 2011

Prof. Dr Ramin Yahyapour, GWDG, Göttingen, Germany

Resource Management in Grid and Cloud Systems

Abstract: Whilst grids have become a common production infrastructure for several scientific research disciplines, cloud computing has gained a broad customer base for mainstream commercial applications. This talk addresses practical and theoretical scheduling problems for grid systems and current work in the area of supporting service-level management for cloud infrastructures. An outlook will be provided on application scenarios for utilising virtualisation technologies in scientific infrastructures.

1 November 2011

Doc. Mgr. Vít Vondrák, Ph.D., VŠB-TU Ostrava, IT4Innovations Centre of Excellence

Development of scalable algorithms for solving highly demanding engineering problems

Abstract: The FETI (Finite Element Tearing and Interconnecting) domain decomposition method, first introduced by Farhat and Roux, has proven to be one of the most successful methods for the parallel solution of linear problems described by elliptic partial differential equations. Its main feature is the decomposition of the domain into non-overlapping subdomains, which are ‘glued together’ using Lagrange multipliers in such a way that, after eliminating the primary variables, the original problem is reduced to a small, relatively well- conditioned quadratic programming problem with a linear constraint, which is then solved iteratively.

The presentation will introduce an efficient massively parallel implementation of our variant of the FETI domain decomposition method, which we call Total FETI, and its efficiency will be demonstrated on complex engineering problems such as those involving material or geometric non-linearities, or optimal design problems. Finally, we will present a completely new variant of the FETI method, which we call H-FETI (Hybrid FETI), which enables parallel implementation across hundreds of thousands of cores.

8 November 2011

Assoc. Prof. RNDr. Petr Sojka, Ph.D., Faculty of Informatics, Masaryk University

The Art of Mathematics Retrieval

Abstract: The design and architecture of MIaS (Math Indexer and Searcher), a system for mathematics retrieval, are presented, and design decisions are discussed. We advocate an approach based on Presentation MathML utilising the similarity of mathematical subformulae. The system was implemented as a maths-aware search engine based on the state-of-the-art Apache Lucene system and is used in The European Digital Mathematics Library – EuDML.

Scalability issues were tested using more than 400,000 arXiv documents containing 158 million mathematical formulae. Almost three billion MathML subformulae were indexed using a Solr-compatible version of Lucene.

15 November 2011

Prof. RNDr. Radim Bělohlávek, DSc., Faculty of Science, Palacký University Olomouc

Formal concept analysis of data with fuzzy attributes: recent developments and related topics

Abstract: Formal concept analysis is a method of data analysis with roots in traditional Port-Royal logic, and applications in various fields including software engineering, information retrieval, homeland security, and psychology. At the core of FCA lies the mathematics and algorithms for relational data, in particular for closure structures, Galois connections, and finite partially ordered sets. In the basic setting, FCA works with binary data. The talk will provide an overview of an extension of FCA to data with fuzzy (graded, ordinal) attributes, the foundations of which have been developed by the speaker and his group over the past ten years. The talk will survey the basic structures underlying FCA for data with fuzzy attributes, the algorithms involved, and the relationships to FCA for data with binary attributes. In addition, connections to some recent topics, such as factor analysis of relational data and the relational model of data over domains with similarities, will be presented.

22 November 2011

Prof. Václav Rajlich, Wayne State University, Detroit

Current trends in the development and teaching of software engineering

Abstract: This lecture reviews the challenges and constraints faced by the lecturer of a software engineering course. It argues that the best introduction to the discipline of software engineering is training in the role of developers within a directed iterative process (DIP), where the most common task is software change (SC). In the course projects, students practise their skills by working on medium-sized open-source software systems, whilst the lecturer fulfils all the other DIP roles. A comprehensive overview of the SC phases – including refactoring, concept identification, impact analysis, unit testing, etc. – forms the core of the course. Finally, the course briefly reviews the rest of the software engineering discipline.

The results show that this course structure provides students with a more realistic experience than traditional software engineering courses. The course has been taught on several occasions at Wayne State University and students have expressed a high level of satisfaction. The resources required for such a course are comparable to those for other computer science courses. A new textbook supporting this approach is introduced [1].

[1] Vaclav Rajlich, Software Engineering: The Current Practice, CRC Press, 2011

29 November 2011

Dr Reinhold Huber-Mörk, Austrian Institute of Technology, Vienna

Automatic coin classification and identification

Abstract: We investigate object recognition and classification in a setting with a large number of classes, as well as the recognition and identification of individual objects of high similarity. Real-world data sets were obtained for the classification and identification tasks. The classification task under consideration involves distinguishing modern coins into several hundred different classes. Identification is investigated for hand-made ancient coins. Intra-class variance due to wear and abrasion, combined with low inter-class variance, makes the classification of modern coins challenging. For ancient coins, the intra-class variance makes the identification task feasible, as the appearance of individual hand-struck coins is unique. We will present methods for coin image classification and identification, along with results for large real-world datasets of modern and ancient coins.

6 December 2011

Prof. Herbert Edelsbrunner, IST Austria, Vienna

Alexander duality for functions

Abstract: Consider a decomposition of the (n+1)-sphere into spaced U and V whose intersection is an n-manifold, M. Alexander duality relates the homology of U to that of V, and using the Mayer-Vietoris exact sequence, we obtain a relation between the homology of M and U. This talk presents extensions of this classical version of Alexander duality to real-valued functions. One of the results is as follows:

Let A be a compact set in R^{n+1}, let its boundary dA be an n-manifold, and let f: R^{n+1} --> R be a smooth function without critical points whose restriction to dA is tame. Then the persistence diagram of f restricted to dA is the disjoint union of the persistence diagram of f restricted to A and the reflection of this diagram.

Joint work with Michael Kerber.

13 December 2011

RNDr. Jan Pomikálek, Ph.D., Faculty of Mathematics, Masaryk University

Doc. PhDr. Karel Pala, CSc., Faculty of Mathematics, Masaryk University

Web corpus in one click

Abstract: Text corpora have a wide range of applications in natural language processing. The web has become a very popular source of data for corpora in recent years. However, there are many challenges associated with creating web corpora, such as web crawling, character encoding detection, language identification, removing junk, and deduplication. We believe we have found suitable solutions to all these problems. We have also developed software tools that make it possible to create a web corpus in a fully automated manner for any language with a reasonable presence on the web and on Wikipedia in particular. Our latest experiments show that for ‘large’ languages (such as English or Spanish), we can collect as many as one billion words of clean text without duplicates in a single day using a single powerful server. We can also easily create corpora for less-resourced languages, such as Tajik, though, of course, at a much slower rate. The automation of web corpus creation can be such that the only input required is the Wikipedia code for the target language. In the presentation, we will describe our processing pipeline, discuss how some of the biggest challenges are addressed, and present our preliminary results.