MN Seminar

Seminars will be held in attendance on Tuesdays in room C10 | 9.01 at 4:15 pm.

16/01/2024
Timo Schürg

Topological data analysis for detecting faults in the mobile network

Topological data analysis uses methods from algebraic topology to generate interesting features for machine learning from the geometry of the data. I would like to present how these methods can be used to detect disturbances in a mobile network. The type of interference is not known beforehand.

12/12/2023
Romana Piat

Numerical modeling of particulate composites

Numerical evaluation of the elastic and electric properties of particle reinforced composite was provided. For understanding the effects of the particle shapes onto the overall linear elastic and electric properties of the two-phase composites the particles with different shapes were modelled using analytical functions. Elastic and elastic properties of the composites with different particle shapes were calculated. Good agreement of the present and known from literature results is achieved.

14/11/2023
Andreas Weinmann

Algorithms for change point detection and piecewise regression

Detecting change points in time series is an active research topic that is closely related to fitting a model to the data between change points. Numerical change point detection algorithms often perform joint estimation of the change points along with the signal chunks. I report our results on this topic based on variational approaches, and methodologically we use dynamic programming approaches in combination with customized solution methods from numerical linear algebra to develop stable solvers.

01/24/2023
Gael Bringout

Magnetic Particle Imaging: An Interdisciplinary Journey

Magnetic Particle Imaging is an imaging modality that was introduced in 2005 by Philips Research in Hamburg, Germany. Its applications are mainly aimed at the medical field, for example to help with diagnoses or to monitor treatments. The research directions include in particular medicine, pharmacy, chemistry, physics, mechanics, electronics and of course mathematics. In the course of this presentation, we will introduce the basics of this technology and highlight the various challenges, especially those that concern mathematicians.

02.05.2023

Prof. Dr. Christoph Raab

Single-frequency diode lasers in industry and research

How can I realize diode lasers to build an optical atomic clock?

This initially rather academic question becomes more and more relevant, because the same techniques become important for industrial applications.

In my talk I will first present some different applications and their requirements. Then I will present the technologies that enable single-frequency and tunable diode lasers.

In doing so, I will explain how a laser diode with a coherence length of millimeters can be turned into a diode laser with a coherence length of 1000km. Finally, I highlight challenges that are still delaying industrial deployment.

06.06.2023
Prof. Dr. Johannes Gregori

Medical image processing and points of connection to OBV.

What are current trends in medical image processing, and how can points of connection to our OBV courses be used?

I will present two research projects currently being conducted with the participation of H-DA, ASPIRE (Alzheimer's disease) and BOSOMSHIELD (breast cancer).

Furthermore, I will show how medical imaging and image processing techniques are linked to OBV lecture content and can be used for lecture design. Examples are CT (industrial and clinical) or MRI (acquisition in frequency space, image reconstruction by 2D or 3D Fourier transform).

07/04/2023
Prof. Dr. Elke Hergenröther

Issues in Computer Vision as Motivation for Cognitive Science and Cognitive Computing

In this talk, we will present questions that we are currently working on. Most of these questions are related to the field of cognitive computing, which is concerned with developing technologies based on artificial intelligence that emulate the human thinking and perception process. On the other hand, we also need the knowledge from the fields of Cognitive Science and Cognitive Computing to be able to evaluate when a process is better than humans or when it meets the requirements of the user.

I would like to discuss these two sides of Cognitive Science and Cognitive Computing in interaction with Computer Vision with you on the basis of some selected case studies.

15.11.2022
Prof. Dr. Thomas März

Model-Based Reconstruction and Regularization for Magnetic Particle Imaging

Abstract: Magnetic Particle Imaging (MPI) is an emerging imaging modality developed by Gleich and Weizenecker in 2005 and is today a very active field of research. In the multivariate MPI setup images are usually reconstructed using a system matrix which is obtained by a time consuming measurement procedure. We approach the reconstruction problem by employing reconstruction formulae which we derive from a mathematical model of the MPI signal encoding. Here, we present a flexible reconstruction algorithm based on the decomposition of the imaging process provided by the model. Its variational formulation incorporates adequate regularization which yields promising reconstruction results.

13.12.2022
Prof. Dr. Christine Bach

Statistical test procedures for the verification of mortality tables

10.01.2023
Prof. Dr. Antje Jahn

Machine learning methods for event time data

Especially in medical research, the endpoints to be analyzed often exhibit so-called censoring. An example underlying this talk is survival after kidney transplantation, which is unknown for all patients not yet deceased at the time of analysis.
Regression models and machine learning methods usually address censoring by adjusting the likelihood or loss function. Alternatively, bagging has been proposed in combination with inverse-probability-of-censoring weighting (IPCW), after which censored observations drop out of the bootstrap samples and therefore no longer require special consideration.
In this presentation, different learning methods are compared using kidney transplant data as an example. It is shown that the IPCW approach leads to biased predictions in the likelihood-free random forests.

Feb 07, 2023
Prof. Dr. Thomas Netzsch

Efficient Programming in the Change of Time

Efficient programming has been an important topic in information technology from the beginning of programming until today.
At the beginning of professional programming in the 20th century, the optimization of organizational processes was an important means to improve resource efficiency. The resource to be optimized was the human being.
Then, for a long time, the skillful use of limited resources such as storage space, computing power and information transfer by developers was central to the feasibility of certain requirements.
The rapid technological progress of the last decades has led to the fact that the resources mentioned are available today in many application areas at low cost and in almost unlimited quantities. In addition, new application areas and technologies, such as climate research or the Internet, have generated an extraordinarily high additional demand for resources.
This globally considerable resource requirement in connection with the broad application of information technology by almost everyone results in the fact that the topic of sustainability has gained considerably in importance and currently the entire life cycle of a software product must be considered in the efficiency analysis. One resource to be optimized is energy consumption.
In this talk, an exemplary overview of old, new and current methods for efficient and resource-saving programming will be given.

03.05.2022
Prof. Dr. Jürgen Frikel (OTH Regensburg)

Microlocal analysis in tomographic reconstruction

14.06.2022
Prof. Dr. Ralf Blendowske

Wave area measurements on eyes in big and small

How does the human eye stay in focus during the growth process and continue to see sharply despite its changes in length? This question has not yet been answered. In a longitudinal study, the eyes of children and adolescents were measured using Hartmann-Shack sensor technology to investigate temporal changes with age. The focus is on the so-called higher-order Zernike coefficients. They provide individual information beyond the values in the eyeglass passport and are being traded as candidates for solving the "sign problem."

07/05/2022
Dr. Maximilan Baust (Nvidea).

Five Ways to Solve the Heat Equation - from Fourier to Deep Learning

Abstract: The heat equation has been playing a crucial role in the invention of many mathematical, and particularly numerical, techniques for solving partial differential equations. Starting with the Fourier transform itself, this talk presents a few milestones along the path of seminal works for solving the heat equation including finite differences, operator splitting techniques, super-time-stepping methods and neural networks. In addition to explaining the mathematical concepts, we will also shed a light on the reason of why modern hardware accelerators, such as GPUs, have gained significant popularity for executing these methods.

02.11.2021
Prof. Dr. Jutta Groos

Bayesian approaches in clinical trials

In many areas of statistics, as well as in clinical trials, classical methods of statistics have prevailed. The efficacy of drugs is usually proven by classical hypothesis tests with a given significance level. Although the Bayesian view of a study situation is not new, it has long been displaced from ̈official perception by the classical (also frequentist) view. Although the two approaches initially appear very different, they very often lead to identical results.

Nevertheless, the Bayesian view offers some advantages, such as a more intuitive interpretability as well as an easy way to consider not only the data of a current study, but also to include earlier findings in a decision. Especially in studies ̈on rare diseases with very small numbers of cases, the inclusion of earlier findings can be useful.

However, if the regulations for clinical trials specify fixed upper limits for the 1st type of error, no power advantage is gained from the use of earlier findings. This talk will explain the alternative Bayesian view using a specific example from cancer research and highlight the problem of no power gain.

07.11.2021
Prof. Dr. Sebastian Döhler

On the control of False Discovery Exceedance in heterogeneous tests.

In statistical analysis of Big Data, often thousands or millions of statistical tests are performed simultaneously. False Discovery Exceedance (FDX) is an important measure of the first kind error that occurs in such situations. The classical methods assume the situation that the distributions of p-values under the null hypotheses are equal. In reality, however, these distributions are often heterogeneous, for example, when discrete tests are present, or when the results of the individual tests are weighted.

In this talk, we present new methods that account for this heterogeneity while controlling for FDX.

Joint work with Etienne Roquain.

11/01/2022
Dr. Raphael Memmesheimer

Service robotics - competitions and applications.

The talk will give experiences and developments from international robotics competitions such as RoboCup, European Robotics League and World Robot Summit. The focus is on autonomous service robots in household environments. These map the environment and
are then able to navigate in it. They recognize objects and people and perform pick-up and drop-off tasks. These application-oriented scenarios regularly benefit from the progress made in research and are becoming increasingly reliable.

Feb 08, 2022
Dr. Lukas Kiefer

Efficient Algorithms for Mumford-Shah and Potts Problems

Data smoothing is often the first step in a data processing chain. Reasons for this can be, for example, the suppression of noise or a simplification of the data. However, the
application of classical smoothing methods can lead to the smoothing out of desired abrupt changes in the data, such as changepoints in time series or edges in image data.

Mumford-Shah and Potts models are among the best known (variational) methods for edge and changepoint preserving smoothing and partitioning of images and time series, respectively. However, their application is not trivial, since difficult non-convex minimization problems have to be tackled. Accordingly, developing new algorithmic approaches to solving Mumford-Shah and Potts models is an active area of research.

In this talk, we consider extensions of the Mumford-Shah and Potts models to higher order models to counter known drawbacks of the conventional first order models. In particular, we highlight the algorithmic requirements that such generalizations entail and present new algorithmic approaches.

Joint work with Martin Storath (Würzburg-Schweinfurt University of Applied Sciences) and Andreas Weinmann (Darmstadt University of Applied Sciences).

20.04.2021
Prof. Dr. Sebastian Döhler

On the analysis of multiple discrete tests: New methods and software

Abstract: The Benjamini-Hochberg procedure and related methods are classical methods that guarantee control of the false discovery rate (FDR) and play an important role in the analysis of high-dimensional data. While these procedures were developed for continuous test statistics, discrete data are often present in applications. In this case, the Benjamini-Hochberg procedure, for example, still guarantees control of the FDR, but it is generally too conservative. Therefore, it is of interest to develop more efficient FDR procedures for discrete data. In this talk, we present such procedures and their implementation in an R package, and illustrate their performance for empirical and simulated data.
Joint work with Etienne Roquain, Guillermo Durand (Sorbonne Universite) and Florian Junge (DISO).

04.05.2021
Prof. Dr. Jan-Philipp Hoffmann

Topological Data Analysis

To apply geometric methods in data analysis, numerically encoded data are understood as a random sample of points on a manifold M embedded in a Euclidean space. If the points are sufficiently close to each other (i.e., if the sample is large enough), the rough appearance of the original structure can be approximately reconstructed by connecting points that have the distance smaller R, for a suitable R>0. Since a suitable R is usually unknown, one studies this connection construct as a function of R as a filtration of simplicial complexes using methods of algebraic topology. The resulting topological invariants can then provide insight into the geometric structure of the data. Finally, we present how the methods of topological data analysis can be used to control deep learning methods, as in artificial neural networks.

01.06.2021
Prof. Dr. Elke Hergenröther

Reality-simulation gap

Convolutional Neural Networks (abbreviation: CNN or ConvNet) are special Deep Learning methods that are suitable for recognizing specific objects in images or videos. They are used in autonomous driving, the recognition of faces and the interpretation of emotions and the like. In industry, too, people want to use these methods for a specific type of problem. To be able to do this, one needs training data in not inconsiderable quantities. Generating this data is usually problematic, which is why one likes to resort to model data. Although, as I will show, current rendering methods can already generate very realistic material properties, there are differences between the model data and the real images. The Reality Simulation Gap is based on these differences. Keeping this gap small is the necessary prerequisite to use simulations instead of real images or videos for training Convolutional Neural Networks. The Reality Simulation Gap is not limited to ConvNets alone. Other Deep Learning methods, such as Deep Reinforcement methods also struggle with the Reality Simulation Gap.

"What is the Reality Simulation Gap and how to try to minimize it?" are the questions that will be discussed in the talk. After the talk, however, you will not only have answers to these two questions, but you will also know what a Convolutional Neural Network is, roughly how it works, what is meant by transfer learning, and why game engines such as Unreal and Unity are becoming increasingly popular even in more traditional industries.

06/29/2021
Prof. Dr. Horst Zisgen

Queuing systems with parallel operators and group arrivals and service

From our everyday life we all know the annoying phenomenon of waiting, be it at the box office, in front of the elevator, in the doctor's waiting room, in the queue of the service hotline, on the highway in a traffic jam or as a skier at the gondola station. However, queues do not only occur in such everyday situations, but also in technical systems, such as web servers, or in the area of production and logistics. Queueing models generally represent the waiting of entities, such as customers, patients, or jobs, in front of an operator station (such as a cash register, doctor, or web server) as a stochastic process and find practical application in many areas of engineering or operations research. However, despite the long tradition of queuing theory and its advances, there are still certain properties of queuing systems that cannot be modeled exactly or at least approximated well. These include, for example, systems at which entities arrive as a group and are then served at multiple operators working in parallel, again as a group but with a different group size. However, it is precisely such group arrivals and group servings that play an important role in many practical problems, such as wafer processing in semiconductor manufacturing, so there is a great need for better modeling of such systems.

In this talk, first, a brief introduction to the state of the art in modeling simplified queuing systems with group arrivals and servicing will be given. Then, a new modeling approach for the approximation of generic systems with group arrivals and service is presented, which offers a significant improvement in approximation quality compared to existing methods, thus enabling practical use in the above mentioned application areas. Finally, the goodness of the modeling is proven by simulation experiments.

17.11.2020
Prof. Dr. Martin Storath

Fast median filtering for phase or orientation data

Abstract: Median filtering is among the most utilized tools for smoothing real-valued data, as it is robust, edge-preserving, value-preserving, and yet can be computed efficiently. For data living on the unit circle, such as phase data or orientation data, a filter with similar properties is desirable. For these data, there is no unique means to define a median; so we discuss various possibilities. The arc distance median turns out to be the only variant which leads to robust, edge-preserving and value-preserving smoothing. However, there are no efficient algorithms for filtering based on the arc distance median. Here, we propose fast algorithms for filtering of signals and images with values on the unit circle based on the arc distance median. For non-quantized data, we develop an algorithm that scales linearly with the filter size. The runtime of our reference implementation is only moderately higher than the Matlab implementation of the classical median filter for real-valued data. For quantized data, we obtain an algorithm of constant complexity w.r.t. the filter size. We demonstrate the performance of our algorithms for real life data sets: phase images from interferometric synthetic aperture radar, planar flow fields from optical flow, and time series of wind directions.

01.12.2020
Prof. Dr. Thomas März

Image Inpainting by Coherence Transport, Algorithm & Theory

Abstract: In image processing the term Image Inpainting refers to the retouching of damaged or undesired portions of a picture. From a mathematical point of view this is a problem of data interpolation, subject to the side condition that the completed image shall look visually plausible. In the last two decades different approaches, using PDE models from order 2 up to 4, have shown that PDE techniques are fruitful here. In this talk we present an Image Inpainting method based on a transport PDE, a PDE of first order. The transport field is constructed from coherence information which we obtain by applying structure tensor analysis. We will demonstrate applications, and discuss the algorithm as well as the analysis of the underlying model.

19.01.2021
Prof. Dr. Christoph Becker

Liquidity Consequences of Central Clearing

Abstract: Central clearing counterparties transform counterparty credit risk into liquidity risk for its clients. The volume of required margin from their members causally influences both liquidity risk premia in the core of the US financial system and the liquidity risk exposure of dealer banks. We discuss implications for financial stability and regulation.

09.02.2021
Prof. Dr. Matthias Will

Research and application areas in the electron microscopy laboratory at the h-da

Abstract: Material variety and packing densities have massively increased in all application areas of technology in the last decades. For example, the application of nanomaterials is nothing unusual anymore. Consequently, there has been an evolution of analytical methods such as electron microscopy from more academic use to industrial analysis tools. The departments MN, MK and CuB have established a joint laboratory with a modern electron microscope. The focus in the selection of the equipment was on the analysis of non-metallic materials and use in research and teaching. The lecture will report on the possible applications and research activities.