BMBF Bundesministerium für Bildung und Forschung

Research-Campus MODAL

G-RIPS - Projekte

2023 sponsors and projects for G-RIPS Berlin

  Project 1: Cray/HPE
    Project 2: 1000shapes GmbH
    Project 3: LBW

Project 1: Cray/HPE

Title: Evaluating the Compiler Optimization Capabilities on Next-Generation Hardware

Sponsor: Cray/HPE

PROJECT DESCRIPTION:
In 2023, a first generation of non-GPU hardware for data-parallel processing combined with compiler-assisted optimization techniques at runtime become available on the market. For the code developer the question arises, how well this concept works in practice, which kind of kernels are best suited for this hardware, and more important, which optimization steps the compiler and runtime system is applying.

The project goal is a characterization of the compiler/runtime capabilities, the robustness of the toolchain, and the relative performance improvements through the automated mechanisms. For the later, a suitable metric is defined within the project.

Requirements for the applicant:
– experiences with code development in C/C++ or FORTRAN
– good knowledge with C/C++ or FORTRAN compiler toolchains (GCC, LLVM, or Cray/HPE)
– basic understanding of using accelerators as offload engines (e.g., GPU, FPGA)
– ideally familiar with source code management (CMake, make, git)
– interest to work with early access hardware and software

Project 2: 1000shapes GmbH

Title: Shape Model Benchmark for Defect Reconstruction

Subtitle: Non-rigid registration of 3D shape priors to sparse or incomplete data

*Sponsor*:  1000shapes GmbH
*Contact*: ZIB (Stefan Zachow <zachow@zib.de>, Marko Leskovar <leskovar@zib.de>)
1000shapes GmbH (Hans Lamecker <hans.lamecker@1000shapes.com>, Dennis Jentsch<dennis.jentsch@1000shapes.com>)

PROJECT DESCRIPTION

*Motivation*
Shape models (including Statistical Shape Models – SSMs) play a crucial role in various applications, such as defect completion, 3D shape estimation from sparse measurement data, or general statistical analysis.  A Statistical Shape Model can be regarded as a population based deformable shape prior that can been generated from many observations. Such a prior does represent the average shape of such an observation including anatomical variation in shape [1]. Using SSMs for a match to given individual data became increasingly popular, e.g., in orthopedics where normally shaped anatomical priors are compared to pathological anatomy of individual patients in order to assess deviations between the two states or to propose corrections that lead to more normal/healthy states [2, 3]. Therefore, SSMs may provide a suitable shape prior to assess anomalies and to guide patient specific surgical reconstruction or individualized implant designs. The driving question is “How much information does a pathological, maybe incomplete anatomical structure give us in view of its unknown native state?”. This is an ill-posed inverse problem which can be solved by using shape priors.

Traditional methodologies have shown great success but also some limitations. Novel methods that utilize non-linear shape spaces [15] or neural network techniques [12, 13, 14], however, are mathematically and computationally challenging and subject of current research. Your task in this project will be to explore where and how those new approaches can make some impact in the above mentioned applications.

*Task*
Within this project we aim to reconstruct patient specific native anatomical regions which are unknown due to pathologies or incomplete data with the help of SSMs. Several SSMs can be investigated for an estimation of a native anatomical approximation by mathematical optimization of a shape fitting process. Therefore, a comparison of different shape modeling frameworks and an evaluation of the respective fitting accuracies need to be undertaken. Besides standard affine registration approaches such as variants of iterative clostest points (ICP) [4] or coherent point drift (CPD) [5], we are primarily interested in non-rigid registration such as non-rigid ICP [6] or Gaussian process morphable models [7], and particularly in SSM-based approaches [1, 2, 8-16]. The following frameworks are potentially suitable for comparison:
•             ZIB/1000shapes SSMs [1, 2]
•             Statismo [8] and Scalismo [9]
*             ShapeWorks [10]
•             Mesh Monk [11]
•             FlowwSSM [12, 13, 14]
•             Deformetrica [15]
•             Morphomatics [16]

*Data*
Different approaches can be tested on a large variety of anatomical shapes, such as pelvic bones, jaw bones, knee bones and many more that have been derived and geometrically reconstructed from tomographic image data. In addition we can also provide facial surface data that have been captured with stereophotogrametry.

*Aims*
A team of 4 G-RIPS students will collaborate on this topic, however, with a challenging aspect of competition.
Besides a common general understanding of SSMs, each student shall make him- or herself familiar with a different approach, where at least one (better two) student(s) shall evaluate neural flow SSMs. The different approaches and the respective results will then be compared and presented. The best approach will be awarded.

*References*
[1] https://opus4.kobv.de/opus4-zib/frontdoor/index/index/docId/7321
[2] https://opus4.kobv.de/opus4-zib/frontdoor/index/index/docId/5693
[3]
[4] https://arxiv.org/abs/2007.07627
[5] https://ieeexplore.ieee.org/abstract/document/5432191
[6] https://www.researchgate.net/publication/200172513_Optimal_Step_Nonrigid_ICP_Algorithms_for_Surface_Registration
[7] https://arxiv.org/abs/1603.07254
[8] https://www.insight-journal.org/browse/publication/871
[9] https://scalismo.org/
[10] https://www.sci.utah.edu/software/shapeworks.html
[11] https://www.nature.com/articles/s41598-019-42533-y
[12] https://opus4.kobv.de/opus4-zib/frontdoor/index/index/docId/8721
[13] https://opus4.kobv.de/opus4-zib/frontdoor/index/index/docId/8855
[14] https://arxiv.org/abs/2211.15601
[15] https://www.deformetrica.org/
[16] https://opus4.kobv.de/opus4-zib/frontdoor/index/index/docId/8544

Project 3: LBW

Title: Electric Bus Scheduling

Sponsor: LBW

PROJECT DESCRIPTION:

Electric Bus Scheduling

Electric busses are used in more and more public transit companies in order to provide emission free public transport. Their operation, however, poses several challenges, in particular, limited range, long charging times, limits on the total amount of energy that is available for charging at any time, and energy prices that fluctuate and/or are related to the peak load.

What is the best strategy to deploy ebusses? Should one prefer large over small batteries, is depot charging better than opportunity charging, and what is a good charging strategy? Such questions can be assessed by using mathematical optimization methods. In fact, the electric bus scheduling problem can be seen as a multicommodity flow problem with additional side constraints on the state of charge of the batteries and the loading facilities.

In this G-RIPS project, we want to work on mathematical optimization approaches to electric bus scheduling. Analyzing and visualizing data, setting up mathematical models, designing und implementing solution approaches, and doing scenario analyses will be the main task in a collaborative work.

The project is supported by LBW Optimization GmbH, a leading supplier of optimization technology for public transit companies, whose solvers are used by hundreds of companies all over the world.

2020 sponsors and projects

1. 1000shapes (Biotechnology)

Company: 1000shapes GmbH is a ZIB spin-off that transfers research in life sciences into products for clinical applications.

Project “Towards an understanding of Anaphylaxis”:

An Anaphylactic shock is a severe allergic reaction which is in most cases caused by food, insect bites or drugs. Individuals experiencing an Anaphylactic shock usually collapse due to a dramatic drop of blood pressure which leads to oxygen under-supply in the cells. If not treated immediately, this can lead to death. Unfortunately, still many reasons for getting an Anaphylactic shock are unknown.

Within this project we will analyze a medical database containing information about Anaphylaxis episodes. The goal is to work towards a better understanding of individual risk factor that can be used for personalized risk estimation of an Anaphylactic shock and the medical conditions that might lead to it. We will use, adjust and extend state-of-the-art machine learning algorithms to extract and understand possible correlations from the given database and will compare to and enrich them with data from literature and other available medical sources.

 

2. Cray Germany GmbH (HPC):

Company: Cray manufactures supercomputer and systems for data storage and analytics. Several Cray supercomputer systems are listed in the TOP500, which ranks the most powerful supercomputers in the world.

Project “Code Analysis for high performance computing”:

Polyhedral compilation encompasses the compilation techniques that rely on the representation of programs, especially those involving nested loops and arrays, thanks to parametric polyhedra or Presburger relations, and that exploit combinatorial and geometrical optimizations on these objects to analyze and optimize the programs. Initially proposed in the context of compilers-parallelizers, it is now used for a wide range of applications, including automatic parallelization, data locality optimizations, memory management optimizations, program verification, communication optimizations, SIMDization, code generation for hardware accelerators, high-level synthesis, etc. There has been experience in using such techniques in static compilers, just-in-time compilers, as well as DSL compilers. The polyhedral research community has a strong academic background, but more and more industry users start to adapt such technologies as well.

Many scientific codes have a strong Fortran code base, and have core regions that are very well suited to polyhedral compilation techniques. However, automatic tools to extract these regions at the source level are lacking. This project will explore integration of the pet (http://pet.gforge.inria.fr/) polyhedral extraction tool into the F18 (https://github.com/flang-compiler/flang/wiki) Fortran compiler, the upcoming high performance Fortran compiler for the LLVM compiler family.

 

3. Deloitte Deutschland (Data Analytics)

Company: Deloitte provides audit, risk advisory, tax, financial advisory, and consulting services to public and private clients spanning multiple industries. Deloitte’s approximately 286,000 professionals are committed to making an impact that matters.

Project “Explainable AI”:

The goal of this project is to assess the interpretability of machine learning techniques in the context of time series forecasts. Based on real sales data, time series patterns need to be analyzed in order to achieve adequate forecasts. Based on both statistical and machine learning approaches, competing forecasting models will be investigated. As a result, the applied approaches will be compared with respect to accuracy, complexity and interpretability.

Additionally, the results will be visualized and current techniques of interpretable Machine Learning will be discussed. You will work together with experienced data scientists and you will learn in this project how to visualize and to communicate results within a heterogeneous team. Furthermore, you will get deep insights into the challenges that practitioners are facing, especially when the achieved solutions to such real-world problems need to be implemented into applied risk management processes.