July 2005

W³odzis³aw Duch

Short update: see list of projects here

My current (2005) research interest include:

0. Creating artificial minds.

So far we are writing projects and looking for money to create Humanized InTerfaces (HITs), integrationg perception, affect, cognition and behavioral control in robotic pets, androids and simulated talking heads for portable devices. This is a large project that will take many years. Proposals have been written to: create a software platform for HIT, test classical and new cognitive architectures in real-time control of android heads, create generative models of emotional processing and recognition, use this knowledge in numerous applications, simulate the process of learning skills, from fully conscious attention to automatization (using the real data from learning how to drive), and make large-scale simulation of the brain functions, based on the CODAM model of John Taylor. This is done mostly in Singapore. An additional project on genuine creativity in the artificial systems is in progress in Toruñ.

We also work on text understanding, annotation, analysis of medical texts, semantic internet and relations ot the brain processes responsible for our language competence. This is done in cooperation with Children's Hospital Research Foundation in Cincinnati, and my students in Poland.

I. Data mining and data understanding methods.

We have developed several logical rule extraction methods, applied them to benchmark problems and to real-world data mining problems. Some of our results are stored as reference in the UCI machine learning repository. A short summary paper describes the main ideas is:

Long term purpose of this research is to construct automatic tools (computer programs) for analysis of the real-world complex data and use this analysis in classification and decision support systems (DSS). Analysis should provide accurate but simple description of the data, first in terms of crisp logical rules, and if this fails using fuzzy rules, and finally if this also fails - for example, if accuracy is too low or the number of rules too large - general neural techniques of classification will be used. Understanding of the structure of real data in terms of simple rules has great importance in many branches of science, technology and especially in medicine.

The methods developed by us for logical rule extraction are based on constrained multilayer perceptron backpropagation networks (MLP2LN method) and their constructive version C-MLP2LN. MLP2LN is based on multilayered perceptrons trained first on the data and subsequently simplified to obtain a skeleton network with 0 or +/-1 weights. It uses the same modification of the error function as the C-MLP2LN method, but the algorithm is not constructive and more than one hidden layer may be used. The C-MLP2LN is based on constructive multi-layered perceptron algorithm, adding one new neuron per class and enforcing logical interpretation by modification of the error function ensuring smooth transition to the 0 and +1 or -1 weights. After this transition only a few relevant inputs are usually left. Dominant rules are recovered from analysis of the first input weights, interpreted as 0 = irrelevant feature, +1 = positive evidence and -1 = negative evidence. New neurons are added, their connections simplified and more specialized rules recovered.
These methods are described in the following papers:

FSM2LN is another rule extraction method that we routinely use in complex cases to generate initial logical rules is based on the probability density estimation neural network called the Feature Space Mapping (FSM). This is a constructive density estimation network that uses separable transfer functions, therefore each node corresponds directly to a fuzzy logical rule with context dependent membership functions, different for each attribute. Traditional membership functions are obtained from the context-dependent membership functions as a special case. Error function includes an additional regularization term that spreads density over all data range in absence of evidence to the contrary, enabling feature selection. Crisp logic is obtained if rectangular functions are used in density estimation or if biradial functions are used (similar to trapezoidal membership functions), allowing for a smooth transition from fuzzy to crisp logic interpretation. The method is described in the following papers:

Our new neural logical rule extraction method is based on a constructive constrained discrete multilayer perceptron network (CD-MLP2LN method), with search techniques used instead of gradient based backpropagation minimization. It is described in:

We also have a decision tree logical rule extraction method based on separability criterion, described in:

We have developed a complete methodology of logical rule extraction from data using C-MLP2LN method for initial generation of rules and global minimization techniques to optimize the rules. An important part of optimization of logical rules is based on assumption that the data is not measured precisely. In effect rules become fuzzy, with specific membership function, and may be optimized using gradient techniques even in complex cases. This methodology is described in papers quoted above and in:

We have made a direct comparison of our C-MLP2LN method with several other methods providing many benchmark results for other experts, for example in papers:

II. Development of neural, machine learning and neurofuzzy systems

We are particularly interested in systems capable of neural-like adaptation and fuzzy expert system reasoning. Our Feature Space Mapping neural model was inspired by results from cognitive science and our understanding of how the neurodynamics of the brain should be approximated to create a system with mind-like behavior.

Some more information are in the neural transfer functions project

III. Development of a general theory integrating machine learning and neural networks.

We have made significant progress working on a theory that integrates many methods from machine learning, neural networks and pattern recognition. The search for optimal model is performed in a space of all possible models that may be created within one framework, opening many qualitatively different optimization channels. Similarity-rules may be created, more general than fuzzy or crisp logic rules. Similarity-based learner software that has been created in my lab will be included in our Ghostminer data-mining package.

See recent talks here, and publications relevant to this topic are (and more here):

  • Duch W, Towards comprehensive foundations of computational intelligence. | PDF file.
    In: W. Duch and J. Mandziuk, Challenges for Computational Intelligence. Springer Studies in Computational Intelligence, Vol. 63, 261-316, 2007.
  • Duch W, A framework for similarity-based classification methods, Intelligent Information Systems VII, Malbork, Poland, 15-19.06.1998, pp. 288-291
  • IV. Applications of neural and machine learning systems

    Theoretical developments result from ambitious applications to prediction, classification and logical rule extraction problems. We have used our rule extraction techniques in a number of applications, including large-scale real world applications in medicine, psychometry, and marketing research. Our data mining methods have discovered knowledge in the form of simplest logical rules for a number of medical and technical datasets. A 670 pages book on applications to psychometry appeared in March 1999 (in polish, since the software is designed for clinical psychologists in Poland). Some papers on this topic include:

    We have developed Intelligent Decision Support Systems for psychometric and medical applications based on the logical rules derived from psychometric and medical tests and plan to commercialize some of them. We also work on cognitive models of the mind collaborating with psychologists.

    V. Other stuff

    My older work on the "Symmetric group graphical approach (SGGA) to the direct configuration interaction method" (Duch W, Karwowski J, Int J Quantum Chem 22 (1982) 783-824; Duch W, Karwowski J (1985) Symmetric group approach to configuration interaction methods. Computer Physics Reports 2:92-170) has not been forgotten! SGGA was implemented by myself and Jacek Karwowski in Munich (1985-87), then another implementation by the Munich group was made (1989-95, Geerd H.F. Diercksen, Norbert Flocke, Erich Schreiner, Peter Graf, Shigeyoshi Yamamoto, with contributions from Jacek Karwowski and Maria Barysz); by Volker Pless, at the Rheinishen Friedrich-Wilhelms-Universitat, Bonn, Germany (1994); by Paul Strodel (supervised by Paul Tavan) at LUM university, Munich, and by Derek Walter at UCLA (1998-2001).

    Robert Harrison at PNL implemented my newer inner projection CI algorithm for full CI. Jarek Meller has finished some work started during his PhD (finished in 1996) and this will be probably my last quantum chemistry paper.

    I have also done some work on foundations of physics in the late 1980s, including EPR correlations and Bell's inequality, going to conferences and thinking about quantum paradoxes. In 1988 I have presented an interesting paradox "Violation of Bell's inequalities in interference experiments". In: Open Problems in Physics, Eds. Kostro L, Posiewnik A, Pykacz J and Zukowski M, (World Scientific, Singapore), pp. 483-486, that prompted some people to work in what is now called entanglement and quantum computing. In 1989 I wrote a paper "Complementarity, Superluminal Telegraph and the Einstein-Podolsky-Rosen Paradox". In this paper I have proposed to measure correlations between pairs of particles in a double Mach-Zehnder experiment. Unfortunately I had not time or energy to push the paper through the refereeing procedure. I did send it to two journals and then gave up. Referees found it "seriously flawed" and "completely wrong". In 1993 R.Y. Chiao, P.G. Kwiat and A.M.Steinberg (Berkeley) described exactly the experiment (Phys Rev A, Sci Am. Aug. 1993) I proposed much earlier, even the figure with the experimental setup differed very little. The time was ripe ...
    I wrote a few other crazy papers on foundations, including one on synchronicity. Although I am not active in this field I have retained deep understanding of quantum mechanics and some interest in quantum computers. The past comes back from time to time.

    In addition to being a head of the Department of Informatics (formerly Department of Computer Methods), Faculty of Physics, Astronomy and Informatics, I have a Research and Development company, DuchSoft, which in collaboration with FQS Poland (a subsidiary of Fujitsu) should market several innovative systems soon: software for data mining, especially psychological evaluation, testing, analysis of questionnaires and cognitive toys for infants.

    In 2001 I become a President of the Executive Board of the company, although I strongly resisted getting involved in business. To my surprise the board was full of important computer science professors ... Unfortunately as many other e-business startups it was closed at the end of 2002. Still it was interesting experience.

    I have worked for short time at various departments, such as applied mathematics (UK), astrophysics (Germany), artificial intelligence (France, Japan), chemistry (USA, Canada, Japan), education (UK), engineering (Japan), physics (Poland , Sweden, USA), psychology (Germany). Between 2000-2006 I was an EU expert in "life sciences", Marie Curie panels, later reviewing larger projects.

    I have written 3 books, co-authored two other books on neural networks and on data mining in application to psychometric problems, and plan to write several books on introduction to cognitive science and AI. I am working towards creation of cognitive science at our university. Because of this highly interdisciplinary project I am collaborating with people in biocybernetics, philosophy, education, psychology and neurobiology departments. About 70 students in cognitive science were accepted in 2009 at NCU.

    I am the co-founder and the head of the scientific editorial board of the Polish Cognitive Science and Media in Education journal.
    I am also on the board of the IEEE Transactions on Neural Networks as books and media editor, (since 2000), Computer Physics Communications (Elsevier, North Holland), Special Editor since 1994, International Journal of Transpersonal Studies (since 2000), and The Journal of Mind and Behavior, as assessing editor (since 2002).

    Full CV, list of papers and current projects is attached to my WWW home page.

    Experience in project organization, fund rising etc.
    My main international grants and collaboration I have been involved in include:
    Coordination of the 1992-95 CEC TEMPUS Office project "Computer Based Education", 3 years; total budget 785.000 ECU (almost 1 milion US$ at that time). Participating institutions: N. Copernicus Univ, Univ. of Cambridge and Univ. of Leeds (UK), Paul Sabatiér Univ. and Univ. de Champagne, Reims (F), Max Planck Inst. for Astrophysics, Technical Univ. Munich (DE). I wrote the application to European Commission, found collaborators abroad, found local people from all 9 faculties of our university, established 7 computational laboratories, and did send/received 140 people (staff and students), negotiating projects with many of their potential supervisors, and talking to hundreds of people before selecting good candidates. This was the only time in my life I had a secretary, and she was in Leeds. After the project was over I thought I will wait until my university will provide some administrative support before embarking on another one. Well, I am still waiting.

    First co-investigator of the European Union COST project "Intelligent software for chemistry", 1995-1998. In collaboration with Max Planck Institute of Astrophysics (Garching b. Munchen, prof. Diercksen, principal investigator), Belfast (Ireland), Bratyslawa (Slovakia), Groningen (Holland), Leeds (UK) and Espoo (Finland).

    A new project, "Breeding creative information societies", has been submitted in April 2001 to the Future Emerging Technologies of the 5th Framework European Union program in collaboration with Heidelberg, Munich, Tartu and Ulster groups but it never got off the ground.

    Another project for Future Emerging Technologies, with 6 groups from the UK, Germany, Italy and Greece, has just been submited in 2004. Artificial Brain Architecture and Cognitive Control Understanding System (ABACCUS), has an extremly ambitous goal of creating a global neural brain simulator for robot control.

    Two large projects have been submitted in 2004/05 in Singapore: Humanized Interfaces (HIT), and DREAM. The first one should create talking heads endowed with natural perception, the second one an android head.

    Projects that will change the world and make each child a genius in may respects are slowly starting in collaboration with Children's Hopsital Research Foundation in Cincinnati, Ohio, USA. No, it is not about genetic manipulation ...

    I have organized the Engineering Applications of Neural Networks (EANN'99), in Warsaw, 1999 and am general co-chair of the International Conference on Neural Networks (Warsaw, Sept 2005).