The brochure of the John von Neumann Institute for Computing is available in English and in German. It can be ordered at the NIC secretariat (nic@fz-juelich.de).

Introduction	Super- computing	Astro- physics	Elementary Particles	Many Particles	Polymers	Chemistry	Environment	Other Fields

    Other Fields of Application

The following examples provide an additional impression of the large variety of issues explored at NIC.

Interdependences in EEG Activities of Different Brain Areas	Jochen Arnhold, Peter Grassberger, NIC Research Group Many-Particle Physics; Klaus Lehnertz, Christian E. Elger, Clinic of Epileptology, University of Bonn
Functional Magnetic Resonance Imaging in Real Time (FIRE)	Stefan Posse, Institute of Medicine, Research Centre Jülich
The Parallel Chess Program "Zugzwang"	Rainer Feldmann, Burkhard Monien, Mathematics/Computer Science Division, University GH of Paderborn
Structural Optimization of Neural Networks by Evolution and Individual Learning	Bernhard Sendhoff, Werner von Seelen, Institut für Neuroinformatik, Ruhr University Bochum

Interdependences in EEG Activities of Different Brain Areas

Synchronization processes are essential in the activity of the brain. In a healthy brain only they permit various functions to be interlinked, such as lip movement, movement of the eyes, and "understanding", when reading aloud. Many diseases, such as epilepsies, involve disturbed synchronization. During an epileptic fit, large areas of the brain generate coherent oscillations. It would be of great interest if these oscillations could be predicted and conclusions about permanent damage could be drawn from the changed capability of (de)synchronization. This is exactly what we intend to achieve by means of a new technique of measuring dependencies between signals originating from various locations in the brain.

The illustrations were obtained from signals derived from a rectangular grid of electrodes placed underneath the skull. Light red regions are on average strongly correlated with the other regions, black ones hardly at all. Each of the four pictures corresponds to a different activity: generating words, listening, reading aloud, and speaking. We can see that various regions are involved to different extents. With patients suffering from epilepsies these patterns could be changed, indicating pathological brain tissue which can be surgically removed.

(Jochen Arnhold, Peter Grassberger, NIC Research Group Many-Particle Physics; Klaus Lehnertz, Christian E. Elger, Clinic of Epileptology, University of Bonn)

Functional Magnetic Resonance Imaging in Real Time (FIRE)

One of the most fascinating applications of fast magnetic resonance (MR) imaging is the mapping of brain functions by measuring changes in blood flow during neural activation. Changes in blood oxygen saturation during neural activation generate magnetic field inhomogeneities which can be detected by MR imaging methods with high temporal resolution (about 100 ms per tomogram).

Currently used image analysis techniques for the detection of brain activation can only be applied after the measurement and are very time-consuming. In contrast, a real-time analysis of image data during the measurement offers a number of advantages (e.g. optimization of the stimulation conditions and quality control) and opens up new neuroscientific applications, e.g. in the field of biofeedback. Within the framework of the Functional Imaging in Real Time (FIRE) project possible applications of functional magnetic resonance imaging in real time are being explored at Research Centre Jülich. In cooperation with Siemens Medical Systems and Algorithmicon, a Siemens Vision 1.5 Tesla whole-body scanner has been modified for real-time measurements. Real-time correlation analysis of image intensity changes has been developed and implemented on Unix workstations and on the CRAY T3E supercomputer at NIC-ZAM. This iterative correlation technique can be limited to a time window of freely selectable width, which is moved on with every new measurement to determine the temporal dynamics of brain activation. Primary sensor activations of the visual, motor and auditory cortex can be detected within a few seconds. In addition, new spectroscopic imaging methods were developed, which increase the detection sensitivity for brain activation and thus even permit a real-time detection of neural activations for the control of individual finger movements. Within the scope of the gigabit project, it is planned to develop remote data transmission and 3D visualization of measurement results for applications in telemedicine.

(Stefan Posse, Institute of Medicine, Research Centre Jülich)

The Parallel Chess Program "Zugzwang"

Chess is the most prominent two-person zero-sum game with complete information, especially since the match between the "Deep Blue" IBM computer and world champion Garry Kasparov. The most widespread solution approach for computer chess is "game tree search", a procedure which is also used to solve scientific and technical problems, for example, in electronic circuit design. The underlying alpha-beta algorithm skilfully uses results from previous parts of the search to speed up subsequent partial searches by proving the search to be superfluous and skipping it. It was long considered to be non-parallelizable. In Paderborn, however, it has been possible to efficiently program this algorithm for parallel computers thus reducing the computing time. Parallelization is essentially based on an efficient dynamic load balancing and information distribution. In July 1998 the Paderborn chess program "Zugzwang" on CRAY T3E-512 played the Lippstadt grandmaster tournament with a large number of entrants and finished second. The illustration shows Zugzwang's (white) game against Luke McShane (black) before the 21st move. Zugzwang played 21.c2-c3 and announced a mate in 18 moves! McShane, the later winner of the tournament, resigned after another three moves.

(Rainer Feldmann, Burkhard Monien, Mathematics/Computer Science Division, University GH of Paderborn)

Structural Optimization of Neural Networks by Evolution and Individual Learning

The principles of natural evolution are used for optimizing neural structures (imitating the information processing in the biological brain). The population-based selection of the best-adapted individuals by competition can be optimally realized on the CRAY T3E computer structure.

The individual growth process of the neural network is coupled to a learning process motivated by biological systems. Since the individual development and evaluation is compute-intensive, the CRAY T3E single-processor performance can be optimally utilized.

Coupling the single processes via a master process realizes competition of the individual structures. Only the best networks can produce offspring for the next generation. The neural structures are thus optimized by a coupled evolution and learning process for particular tasks (e.g. prediction of time series, control tasks).

(Bernhard Sendhoff, Werner von Seelen, Institut für Neuroinformatik, Ruhr University Bochum)

Introduction	Super- computing	Astro- physics	Elementary Particles	Many Particles	Polymers	Chemistry	Environment	Other Fields