Malte Prieß

selected publications & proposals

1. KI_INVOICE

   On the need of further evaluation metrics for automated visual document analysis in business environments / submitted

   Mehra, A.,  Prieß, M., and Himstedt, M.

   2024

   Abs Bib

   This paper provides a comprehensive evaluation of the open-source model LayoutLM in a business context focussing on the automated visual recognition of named entities in German invoice documents. We demonstrate that conventional classification metrics are insufficient for evaluating the applicability of machine learning models in practise. The applicability of these models crucially relies on the level of automation, i.e. minimizing human-in-the-loop interventions. To assess the problem we introduce a novel metric, \textitDocument Integrity Precision (DIP), that – to the best of our knowledge – has not been introduced in this context. DIP describes how many documents of the test dataset require manual interventions. It enables AI researchers and software developers an in-depth investigation of the level of process automation in business software. In order to validate DIP, we conducte experiments with LayoutLM to analyse impact and relevance of underlying metrics in different training settings. Our results show that existing metrics barely change for isolated model impairments whereas DIP halves which indicates that the model requires substantial human interventions in deployment. The larger the set of entities being predicted the less sensitive are conventional metrics entailing poor automation quality. DIP, in contrast, remains as single value to be interpreted for entire entity sets.

   @article{Mehra2024,
     author = {Mehra, A. and Prie{\ss}, M. and Himstedt, M.},
     title = {On the need of further evaluation metrics for automated visual document analysis in business environments / submitted},
     year = {2024},
     series = {Proceedings of the 26th International Conference on Business Informatics (CBI 2024)},
     keywords = {document analysis, LayoutLM, NLP, natural language processing, transformer, invoice recognition, token classification, multi-modal},
   }

2. KI_Ethik

   Ethische Rahmenbedingungen beim Einsatz von KI im Projektgeschäft – Gefahren und Chancen der KI

   Prechel, K., and Prieß, M.

   2024

   Abs

   In diesem Kapitel wird der ethische Rahmen beim Einsatz von Künstlicher Intelligenz (KI) im Projektgeschäft beleuchtet. Mit dem Fortschritt der KI ergeben sich sowohl Chancen als auch Gefahren im Projektmanagement. Während KI die Effizienz und Präzision von Projekten steigern kann, birgt sie auch ethische Herausforderungen, insbesondere in Bezug auf Fairness, Datenschutz und Transparenz. Kerstin Prechel und Malte Prieß diskutieren die Relevanz ethischer Überlegungen im Projektmanagement und stellen Fragen zur Integration von KI in ethisch verantwortungsvolle Weise. Das Kapitel betont die Bedeutung von Transparenz, Verantwortung und Verifizierung im Umgang mit KI und bietet einen Einblick in die ethischen Aspekte des Einsatzes von KI im Projektgeschäft.

3. KI_Bildung

   KI lehren, KI lernen an der DHSH – Potentiale für die Bildungslandschaft

   Prieß, M., and Prechel, K.

   2023

   Bib

   @incollection{Priess2023,
     author = {Prie{\ss}, M. and Prechel, K.},
     booktitle = {INFORMATIK 2023 -- Designing Futures: Zukünfte gestalten},
     publisher = {Gesellschaft für Informatik e.V.},
     title = {{K}I lehren, {K}I lernen an der {DHSH} -- {P}otentiale für die {B}ildungslandschaft},
     year = {2023},
     address = {Bonn},
     isbn = {978-3-88579-731-9},
     pages = {457--460},
     doi = {10.18420/inf2023_56},
     pissn = {1617-5468},
   }

4. KI_Campus

   Fellowship proposal "‘KI lehren, KI lernen an der DHSH"’, KI-Campus (Stifterverband für die Deutsche Wissenschaft e.V.) / accepted

   Prieß, M.

   2022

   Abs Bib

   Mit dem Fellow-Projekt möchte ich Lernressourcen und didaktische Konzepte zum Themenkomplex KI an der Duale Hochschule Schleswig-Holstein (DHSH) ausbauen und zur Digitalisierung an der DHSH beitragen, indem Lehrende den Umgang mit digitalen Lehr-Formaten und Studierende unterstützt durch KI-Campus Ressourcen KI-Einsatzkontexte erlernen. Hierzu beabsichtige ich, im Rahmen eines Moduls zum Thema Data Science und Maschinelles Lernen Ressourcen des KI-Campus im Sinne eines Blended-Learnings zu nutzen. Grundlegende Verfahren des Maschinellen Lernens, eine Einführung in künstliche neuronale Netze sowie einen Ausblick auf den Bereich Deep Learning sollen abgedeckt werden. Weiterhin möchte ich für eine gesellschaftliche Folgenabwägung und kritische Reflektion der eingesetzten Technologien über entsprechende Kurse sensibilisieren.

   @article{Priess2022,
     author = {Prie{\ss}, M.},
     title = {Fellowship proposal "`KI lehren, KI lernen an der DHSH"', KI-Campus (Stifterverband für die Deutsche Wissenschaft e.V.) / accepted},
     year = {2022},
     cite-key = {prpr2022a},
     note = {accepted},
   }

5. KI_Transf

   Evaluierung des Dense Passage Retrievals zur Dokumentensuche in Behörden im Vergleich zu BM25

   Mehra, A.,  Prieß, M., and Peters, T. V.

   2022

   Abs Bib

   Die Suche nach Dokumenten ist eine der Aufgaben, die Maschinen am häufigsten ausführen. Täglich genutzte Suchmaschinen suchen nach Dokumenten, um das Informationsbedürfnis eines Anfragenden zu erfüllen. Aktuelle Forschungsergebnisse zeigen, dass der Bewertungsansatz des Dense Passage Retrieval (DPR) – welcher die zu suchenden Dokumente mittels tiefer neuronaler Netze in Dense-Vektoren umwandelt – in englischer ebenso wie in deutscher Sprache bessere Ergebnisse liefert als aktuell implementierte Algorithmen wie der BM25-Algorithmus. Die Dokumentensuche in deutschen Behörden stellt eine besondere Aufgabe dar, da in diesen komplexe behördenspezifische Fachausdrücke verwendet werden. Im Rahmen dieser Arbeit wurde für die Dokumentensuche zu Corona-Regelungen in der öffentlichen Verwaltung ein DPR mit einem öffentlichen deutschsprachigen Datensatz trainiert und mit dem BM25-Algorithmus von Elasticsearch anhand der Relevanz als Gütemaß verglichen. Die Ergebnisse zeigen, dass das DPR mehr Dokumente mit gesuchten Informationen findet, Synonyme in Suchanfragen berücksichtigen kann und damit das Potenzial zur Verbesserung der Dokumentensuche in Behördenanwendungen hat.

   @inbook{Mehra2022a,
     author = {Mehra, A. and Prie{\ss}, M. and Peters, T. V.},
     pages = {353--362},
     title = {{E}valuierung des {D}ense {P}assage {R}etrievals zur {D}okumentensuche in {B}ehörden im {V}ergleich zu {BM}25},
     year = {2022},
     volume = {AKWI-Tagungsband zur 35. AKWI-Jahrestagung},
     cite-key = {mehra2022},
     doi = {https://doi.org/10.30844/AKWI_2022_26},
     journal = {Monographien},
     keywords = {Dense Passage Retrieval, BM25, Natürliche Sprachverarbeitung, Künstliche Intelligenz, Neuronale Netze, Information Retrieval, Dokumentensuche},
   }

6. SBO_OneShot

   Surrogate-Based and One-Shot Optimization Methods for PDE-Constrained Problems with an Application in Climate Models

   Slawig, T.,  Prieß, M., and Kratzenstein, C.

   In Solving Computationally Expensive Engineering Problems 2014

   Abs Bib

   We discuss PDE-constrained optimization problems with iterative state solvers. As typical and challenging example, we present an application in climate research, namely a parameter optimization problem for a marine ecosystem model. Therein, a periodic state is obtained via a slowly convergent fixed-point type iteration. We recall the algorithm that results from a direct or black-box optimization of such kind of problems, and discuss ways to obtain derivative information to use in gradient-based methods. Then we describe two optimization approaches, the One-shot and the Surrogate-based Optimization method. Both methods aim to reduce the high computational effort caused by the slow state iteration. The idea of the One-shot approach is to construct a combined iteration for state, adjoint and parameters, thus avoiding expensive forward and reverse computations of a standard adjoint method. In the Surrogate-based Optimization method, the original model is replaced by a surrogate which is here based on a truncated iteration with fewer steps. We compare both approaches, provide implementation details for the presented application, and give some numerical results.

   @inproceedings{Slawig2014a,
     author = {Slawig, T. and Prie{\ss}, M. and Kratzenstein, C.},
     booktitle = {Solving Computationally Expensive Engineering Problems},
     title = {Surrogate-Based and One-Shot Optimization Methods for PDE-Constrained Problems with an Application in Climate Models},
     year = {2014},
     editor = {Koziel, S. and Leifsson, L. and Yang, X.S.},
     pages = {1--24},
     publisher = {Springer},
     series = {Springer Proceedings in Mathematics \& Statistics},
     volume = {97},
     keywords = {optimization, climate model, marine ecosystem model, one-shot method, surrogate-based optimization},
   }

7. SBO

   Accelerated Parameter Identification in a 3D Marine Ecosystem Model Using Surrogate-Based Optimization

   Prieß, M., Piwonski, J., Koziel, S., Oschlies, A., and Slawig, T.

   Ocean Modelling 2013

   Abs Bib

   We present the application of the Surrogate-based Optimization (SBO) method on a parameter identification problem for a 3-D biogeochemical model. SBO is a method for acceleration of optimization processes when the underlying model itself is of very high computational complexity. In these cases, coupled simulation runs require large amounts of computer time, where optimization runs may become unfeasible even with high-performance hardware. As a consequence, the key idea of SBO is to replace the original and computationally expensive (high-fidelity) model by a so-called surrogate, which is created from a less accurate but computationally cheaper (low-fidelity) model and a suitable correction approach to increase its accuracy. To date, the SBO approach has been widely and successfully used in engineering applications and also for parameter identification in a 1-D marine ecosystem model of NPZD type. In this paper, we apply the approach onto a two-component biogeochemical model. The model is spun-up into a steady seasonal cycle via the Transport Matrix Approach. The low-fidelity model we use consists of a reduced number of spin-up iterations (several decades instead of millennia used for the original model). A multiplicative correction operator is further exploited to extrapolate the rather inaccurate low-fidelity model onto the original one. This corrected model builds our surrogate. We validate this SBO method by twin-experiments that use synthetic observations generated by the original model. We motivate our choice of the low-fidelity model and the multiplicative correction and discuss the computational advantage of SBO in comparison to an expensive parameter optimization in the context of the high-fidelity model. The proposed SBO technique is shown to yield a solution close to the target at a significant gain of computational efficiency. Without further regularization techniques, the method is able to identify most model parameters. The method is simple to implement and presents a promising and pragmatic tool to calibrate biogeochemical models in a global three-dimensional setting.

   @article{Priess2013a,
     author = {Prie{\ss}, M. and Piwonski, J. and Koziel, S. and Oschlies, A. and Slawig, T.},
     journal = {Ocean Modelling},
     title = {Accelerated Parameter Identification in a 3D Marine Ecosystem Model Using Surrogate-Based Optimization},
     year = {2013},
     pages = {22–36},
     volume = {68},
     keywords = {marine biogeochemical model, parameter identification, model calibration, surrogate-based optimization, rsponse correction, low-fidelity model},
   }

8. GravWave

   Subtraction-Noise Projection in Gravitational-Wave Detector Networks

   Harms, J., Mahrdt, C., Otto, C.,  Prieß, M., and Danzmann, K.

   Physical Review D 2008

   Abs Bib

   In this paper, we present a successful implementation of a subtraction-noise projection method into a simple, simulated data analysis pipeline of a gravitational-wave search. We investigate the problem to reveal a weak stochastic background signal which is covered by a strong foreground of compact-binary coalescences. The foreground, which is estimated by matched filters, has to be subtracted from the data. Even an optimal analysis of foreground signals will leave subtraction noise due to estimation errors of template parameters which may corrupt the measurement of the background signal. The subtraction noise can be removed by a noise projection. We apply our analysis pipeline to the proposed future-generation space-borne Big Bang Observer mission which seeks for a stochastic background of primordial gravitational waves in the frequency range ∼0.1 Hz-1 Hz covered by a foreground of black-hole and neutron-star binaries. Our analysis is based on a simulation code which provides a dynamical model of a time-delay interferometer network. It generates the data as time series and incorporates the analysis pipeline together with the noise projection. Our results confirm previous ad hoc predictions which say that the Big Bang Observer will be sensitive to backgrounds with fractional energy densities below Ω=1E-16.

   @article{Harms2008,
     author = {Harms, J. and Mahrdt, C. and Otto, C. and Prie{\ss}, M. and Danzmann, K.},
     journal = {Physical Review D},
     title = {Subtraction-Noise Projection in Gravitational-Wave Detector Networks},
     year = {2008},
     number = {12},
     volume = {77},
   }