# Seminare 2010

## Seminar 1: Overview of Basel HP2C Projects

Profs. O. Schenk, M. Liebendörfer, S. Goedecker, Uni Basel

No abstract available.

Web: http://www.hp2c.ch/projects/

———————————————————————————————————————————————

## Seminar 2: Toward Brain Computer Interfacing: Algorithms for online Differentiation of Neuroelectric Activities

Prof. K.-R. Müller, TU Berlin (D)

No abstract available.

———————————————————————————————————————————————

## Seminar 3: Meta-Analysis of Published 16S rRNA Data to Connect Sequenced Genomes to their Environmental Contexts

Prof. C von Mehring, University of Zurich (CH)

No abstract available.

———————————————————————————————————————————————

## Seminar 4: Neural Network Potential-Energy Surfaces for Atomistic Simulations of Condensed Systems

Prof. J. Behler, Theoretical Chemistry, Ruhr-University Bochum (D)

Electronic structure methods like density-functional theory (DFT) provide an accurate description of many condensed systems like solids, surfaces and interfaces. Still, the high computational costs of DFT calculations severely limit the size of the systems that can be studied. Since the reliability of the results obtained in theoretical simulations strongly depends on the quality of the employed interatomic potentials, the development of more efficient potentials is an active field of research.

Recently, it has been shown that artificial Neural Networks (NN), which have been used to describe small molecular systems for about a decade, can provide accurate representations of high-dimensional potential-energy surfaces for condensed systems. These NN potentials are based on electronic structure calculations, but once constructed, they can be evaluated several orders of magnitude faster, while the accuracy is essentially maintained. The functional form of the neural network provides access to analytic forces, thus enabling a routine application of NN potentials in molecular dynamics simulations of large systems.

The capabilities of the method are demonstrated for several benchmark systems including clusters and solids of various materials. We show that the structural and energetic properties provided by the NN potential are in excellent agreement with reference DFT calculations.

———————————————————————————————————————————————

## Seminar 5: Where in Data is the Relevant Information? Model Validation by Information Theory

Prof. J. M. Buhmann, Dept. of Computer Science, ETH Zurich (CH)

"A picture is worth a thousand words!" But how to find this knowledge in images? Information society is drowning in data but starving for knowledge. Computational science enables us to build and analyse models of a complexity far beyond what we can understand without computers. Our vastly increased ability to generate very complex computational models shifts the focus of the scientific endeavor to model selection and validation. We advocate an information theoretic perspective where the uncertainty in the measurements quantizes the solution space of the underlying modeling problem, thereby adaptively regularizing the degrees of freedom. A pattern recognition model, which can tolerate a higher level of fluctuations in the measurements than alternative models, is considered to be superior provided that the solution is equally informative. The optimal tradeoff between "informativeness" and "robustness" is quantified by the approximation capacity of the selected model.

Empirical evidence for this model selection concept is provided by cluster validation in computer security, i.e., multilabel clustering of Boolean data for role based access control, but also in high dimensional Gaussian mixture models and the analysis of microarray data. Furthermore, the approximation capacity of the SVD cost function

suggests a cutoff value for the SVD spectrum.———————————————————————————————————————————————

## Seminar 6: Euler, Ritz, Galerkin, Courant - on the Way to Modern Computing

Prof. M. J. Gander, University of Geneva (CH)

The finite element method has become indispensable for the numerical simulation of partial differential equations. But where does this method come from? How was it invented? I will show in my talk how everything started with Euler and Lagrange, and their discrete and continuous formulations of variational problems, which led to the highway of variational calculus. We will then see the fundamental contributions of the Swiss physicist and mathematician Ritz in detail, and his method to compute Chladni figures. The development went further on a detour to Russia, to Timoshenko, Bubnov and Galerkin, who immediately realized the importance of Ritz' method, and used it to solve hard problems in science and engineering. The western world in contrast was at that time more interested in existence and uniqueness proofs around Hilbert and Courant. The value of Ritz' invention was only recognized much later by Courant, who presented the first finite element calculation we were able to find in an address to the AMS. The name Finite Element Method was finally coined by Ray Clough and collaborators at Boeing. The mathematical development of the finite element method was then however just to begin.———————————————————————————————————————————————

## Seminar 7: Image Processing for the High-Resolution 3D Structures Reconstruction of Membrane Proteins from Transmission Electron Microscope Images

Prof. H. Stahlberg, Center for Cellular Imaging and Nano Analytics, Uni Basel

Membrane Protein structure determination is a challenging task of high importance for medical research. We record high-resolution transmission electron microscopy images of two-dimensionally crystallized membrane proteins, which are imaged at various sample tilt angles. The resulting images have an awful signal to noise ratio of sometimes below one percent, so that to the the bare eye the images usually look like featureless fog. In addition, the images are affected by the microscopes point spread function, which is non-linear, varies throughout even single images, and has several parameters of uncertainty.

We develop image processing software for the 3D protein structure determination from such images. Involved algorithms include crystal lattice unbending, Fourier backprojection methods, maximum likelihood approaches, and projective constraint optimization. Our software is available at http://2dx.org. This software and its underlying algorithms will be presented.