Kolloquium Wintersemester 2018/19

23.10.2018

06.11.2018

Neuronale Netze und Machine Learning - Einsatzgebiete in Industrie und Wirtschaft

Uwe Westerhoff
_{saracus consulting GmbH, Münster}

Das Themenfeld Machine Learning, Neuronale Netze und Künstliche Intelligenz hat insbesondere in den letzten kanpp zehn Jahren eine enorme Aufmerksamkeitssteigerung erfahren und dies sowohl im wissenschaftlichen Bereich als auch auf Seiten der Industrie. Inzwischen gibt es kaum eine Branche, welche nicht zumindest von sich behauptet, Verfahren der KI einzusetzen. Im Rahmen dieses Vortrags werden wir einen Blick hinter die Kulissen werfen.  Der erste Teil liefert einen Überblick über verschiedene Neuronale Netze und weitere Algorithmen. Außerdem werden wir uns zumindest kurz anschauen, wie Neuronale Netze letztlich auf Ebene der einzelnen Rechenoperationen arbeiten und trotz vergleichsweise einfacher Einzeloperationen ihre beeindruckende Leistungsfährigkeit erreichen.
Im zweiten Teil stellen wir einige Anwendungsfälle aus Industrie und Wirtschaft vor und beschreiben die jeweils eingesetzten Verfahren.

Einladender:  O. Kamps

Bayesian Statistical Modeling for the Natural and Social Sciences

Paul Buerkner
_{Fachbereich Psychologie, WWU}

In recent years, Bayesian Statistics has gained a lot of popularity in the Natural and Social Sciences. Not only does it offer a consistent framework for probabilistic models, but it also provides researchers with an enormous modeling flexibility. It is this flexibility that allows creating statistical models tailored to the requirements of data, not the other way around. In this talk, I will present some general principles that guide the development, estimation, and interpretation of Bayesian statistical models. Among other things, I will talk about the choice of appropriate response distributions, the modeling of non-linear relationships as well as approaches to model comparisons in the Bayesian framework.

Einladender:  O. Kamps

Fluctuation response and large signal stability in power grids

Dr. Frank Hellmann
_{Potsdam Institute for Climate Impact Research, Potsdam}

In the complex network study of power grids, their dynamics are typically modeled as a Kuramoto model with inertia. In order to simplify the analysis it is often the case that losses are neglected, leading to symmetric effective network Laplacian in the stability and fluctuation analysis. I will present new results and analysis that show that losses, which appear as a phase shift parameters as in the Kuramoto-Sakaguchi type inertial oscillator models, dramatically alter the nonlinear and frequency characteristics of these systems.

Einladender:  O. Kamps

Solvent mediated interactions between nanostructures in suspension and the influence of solvophobicity

Andy Archer
_Loughborough

When pairs of objects immersed in a liquid medium encounter one another there is, of course, a direct interaction between then, but there is also a solvent mediated interaction, which depends on the size difference between the objects and the solvent molecules and many other factors. This interaction can be attractive, repulsive or even both, at different separations and distances between the objects. If the forces between the objects are strongly attractive, then this leads to aggregation of the particles, so controlling carefully these forces are important in maintaining the particles in solution. When the surface of the objects are solvophobic, or have solvophobic patches, then this can lead to strong attractive forces. For example, hydrophobic forces such as these are important in determining protein folding in water and other molecular ordering in living systems. I will give an overview of some of the physics of solvent mediated interactions and show how such potentials can be calculated using classical density functional theory (DFT), focussing in particular on the influence of solvophobicity.

Einladender: Prof. U. Thiele

Machine Learning in Psychiatry - Promises, Problems, and Perspectives

PD. Dr. Tim Hahn
_{Klinik für Psychiatrie und Psychotherapie, WWU}

While our knowledge regarding the biological bases of psychiatric disorders has expanded massively in the last two decades, none of these findings have yet been translated into concrete clinical applications. One reason for this is that commonly conducted statistical analyses – while allowing for inference on the group level – do not enable predictions on the level of the individual. Another reason is the multifaceted nature of mental disorders which necessitates the integration of multiple sources of information when aiming to comprehensively represent relevant patient characteristics. Against this background, methods from the field of machine learning and multivariate pattern recognition have recently gained increasing attention, especially in the area of neuroimaging. The presentation gives an overview of the most recent developments in the emerging field of Predictive Analytics in Mental Health Research, providing concrete examples based on functional and structural Magnetic Resonance Imaging assessments, proteomic data as well as large-scale psychometric measures. Specifically, we will detail best practive project planning and common pitfalls when constructing models 1) predicting therapeutic response, 2) supporting differential diagnoses and 3) optimizing risk detection in the context of prevention and patient management. Finally, we will suggest specific guidelines for Deep Learning and Big Data approaches in multi-center research with the goal of providing the clinician with easy-to-use predictive algorithms and databases enabling early diagnosis and, ultimately, more efficient and successful treatment.

Einladender: Dr. O. Kamps

Formation and Spatial Localization of Phase Field Quasicrystals

Priya Subramanian
_Leeds

The dynamics of many physical systems often evolve to asymptotic states that exhibit spatial and temporal variations in their properties such as density, temperature, etc. Regular patterns such as graph paper and honeycombs look the same when moved by a basic unit and/or rotated by certain special angles. They possess both translational and rotational symmetries giving rise to discrete spatial Fourier transforms. In contrast, an aperiodic crystal displays long range order but no periodicity. Such quasicrystals lack the lattice symmetries of regular crystals, yet have discrete Fourier spectra. We look to understand the minimal mechanism which promotes the formation of such quasicrystalline structures arising in diverse soft matter systems such as dendritic-, star-, and block co-polymers using a phase field crystal model. Direct numerical simulations combined with weakly nonlinear analysis highlight the parameter values where the quasicrystals are the global minimum energy state and help determine the phase diagram. By locating parameter values where multiple patterned states possess the same free energy (Maxwell points), we obtain states where a patch of one type of pattern (for example, a quasicrystal) is present in the background of another (for example, the homogeneous liquid state) in the form of spatially localized dodecagonal (in 2D) and icosahedral (in 3D) quasicrystals. In two dimensions, we compute several families of spatially localized quasicrystals with dodecagonal structure and investigate their properties as a function of the system parameters. The presence of such metastable localized quasicrystals is significant as they affect the dynamics of the soft matter crystallization process.

Einladender: Prof. U. Thiele

Intermittent behavior across scales in biology

Fernando Peruani
_{Université de Nice Sophia Antipolis}

Intermittent behavior is observed in biological systems at all scales, from bacterial systems to sheep herds. First, I will discuss how Escherichia coli explores surfaces by alternating stop and moving phases. Specifically, I will show that a stochastic three behavioral state model is consistent with the empirical data. The model reveals that the stop frequency of bacteria is tuned at the optimal value that maximizes the diffusion coefficient. These results provide a new perspective on how evolution may have reshaped bacterial motility apparatus. Intermittent motion is also observed in sheep, where again a stochastic three behavioral state model provides a quantitative understanding of the empirical data. However, in sheep, individual transition rates depend on the behavioral state of other individuals and collective behaviors emerge. Specifically, I will show that small sheep herds display highly synchronized intermittent collective motion, with the herd behaving as a self-excitable system.

Einladender: Prof. U. Thiele

Computational Social Science

Jun.-Prof. Dr. Annie Waldherr
_{Institut für Kommunikationswissenschaft, WWU Münster}

Computational social science (CSS) is a growing interdisciplinary approach to study social processes using big data, advanced data analytics, and computer simulation. Besides just scaling traditional research approaches to massive data sets, some strands of CSS - such as network science and social simulation - also stress the interconnectedness and nonlinearity of social processes, embracing ideas from complexity science. After giving a broad overview on the main areas of current CSS, I will present an example of studying the public sphere as a complex system with agent-based modeling. In terms of complexity theory, the public sphere can be conceptualized as an arena where heterogeneous, autonomous and adaptive agents interact and self-organize. Multiple feedback mechanisms lead to the emergence of non-linear macro patterns such as news waves or opinion spirals. I show how the public sphere can be modeled as an agent-based system, and how the social simulation approach helps to develop theories of the public sphere, specifically by offering proof of concepts of existing theories, finding underlying mechanisms able to generate observed empirical patterns, and exploring possible futures.

Einladender: Dr. O. Kamps