KonKIS 24 - Konferenz der deutschen KI-Servicezentren 2024 (Conference of the German AI Service Centers 2024)

Europe/Berlin
Göttingen, Alte Mensa

Göttingen, Alte Mensa

Wilhelmsplatz 3 3073 Göttingen
Description

🇩🇪   Die KonKIS 24 war ein voller Erfolg. Wir bedanken uns bei allen Teilnehmenden!

 

Lesen Sie unseren Bericht

Bilder der Veranstaltung gibt es in der Bildergallerie.

Das Programm sowie weiterführende Informationen wie Abstracts und Folien (soweit hinterlegt) finden Sie hier

 

 

🇬🇧   KonKIS 24 was a complete success. We thank all participants!

 

Read our conference report

Pictures of the event you find in the gallery.

The programme and further information like abstracts and slides (as far as they have been provided) you find here

 

Deutsche KI-Servicenzentren.

  • Wednesday, September 18
    • 11:00 AM
      Registrierung, Mittagssnack / Registration, Lunch

      Registrierung: unteres Foyer, Mittagssnack: oberes Foyer
      Registration: lower foyer, lunch: upper foyer

    • 🇩🇪 Eröffnung / Opening Session Adam-von-Trott-Saal

      Adam-von-Trott-Saal

      • 1
        🇩🇪 Begrüßung
        Speaker: Prof. Julian Kunkel (GWDG / Georg-August-Universität Göttingen)
      • 2
        🇩🇪 Grußwort
        Speaker: Prof. Metin Tolan (Georg-August-Universität Göttingen)
      • 3
        🇩🇪 Grußwort
        Speaker: Prof. Volker Epping (Leibniz Universität Hannover)
      • 4
        🇩🇪 Impulsvortrag Ministerium

        Über die Sprecherin

        Seit Dezember 2023 ist Dr. Tina Klüwer Leiterin der Abteilung 5 „Forschung für technologische Souveränität und Innovationen“ im Bundesministerium für Bildung und Forschung.
        Zuvor leitete sie als Direktorin das K.I.E.Z. – das Künstliche Intelligenz Entrepreneurship Zentrum. K.I.E.Z. ist eine Initiative von Science & Startups, dem Verbund der Startup Services der Berliner Universitäten und der Charité Universitätsmedizin. Vor dem Start von K.I.E.Z. war Frau Klüwer Gründerin und Geschäftsführerin der parlamind GmbH, einem Unternehmen für Automatisierung im Kundenservice mittels Künstlicher Intelligenz und Sprachverarbeitung.
        Sie war als Wissenschaftlerin 10 Jahre am Deutschen Forschungszentrum für Künstliche Intelligenz (DFKI) und verschiedenen Universitäten tätig und promovierte in Computerlinguistik an der Universität des Saarlandes.
        Dr. Tina Klüwer ist eine anerkannte Expertin, Managerin und technische Botschafterin für das Thema Künstliche Intelligenz und dessen Umsetzung in der Wirtschaft.

        Speaker: Dr Tina Klüwer (Bundesministerium für Bildung und Forschung)
      • 5
        🇩🇪 Impulsvortrag Wirtschaft. Die Zukunft gestalten: KI als Treiber für Unternehmensinnovation

        In diesem Vortrag präsentiere ich den Einsatz von Künstlicher Intelligenz in der Produktion bei Sartorius, welche mittels Bildanalyse-Techniken die Produktqualität überwacht und zeitnah Fehler automatisch erkennt und klassifiziert. Weiterhin demonstriere ich, wie generative AI-Assistenztools spezifische Anfragen zu Sartorius-Produkten schnell und effizient beantworten und so erheblichen manuellen Aufwand einsparen. Abschließend beleuchte ich die notwendigen Rahmenbedingungen und diskutiere die Herausforderungen, die sich bei der Implementierung und Skalierung von KI-Technologien im Unternehmenskontext ergeben.

        Speaker: Dr Jiayi Yang (Sartorius)
      • 6
        🇩🇪 Keynote KI & Gesellschaft. Vertrauenswürdige Künstliche Intelligenz? Ethik in Design, Entwicklung und Anwendung von KI

        Vertrauenswürdige Künstliche Intelligenz? Ethik in Design, Entwicklung und Anwendung von KI
        (Trustworthy artificial intelligence? Ethics in the design, development and application of AI)

        Spätestens seit der Veröffentlichung von ChatGPT ist Künstliche Intelligenz (KI) in aller Munde. Allerdings durchdringt KI schon viel länger unseren Alltag in vielen Bereichen des öffentlichen und privaten Lebens. Dies reicht von Suchmaschinen und Empfehlungssoftware über Wettervorhersagen und Mustererkennung in der Krebsdiagnostik bis hin zur Verwendung von Software, welche Entscheidungen im Sozial-, Bildungs- oder Justizwesen unterstützen soll. Dieser vielfältige Einsatz von KI bringt viele Vorteile mit sich, wirft aber auch zahlreiche ethische Fragen auf. In meinem Vortrag werde ich zunächst erläutern, welche Rolle Ethik in der Technikentwicklung spielt und was hinter dem Schlagwort Vertrauenswürdige KI steckt. Kann man KI überhaupt vertrauen - und sollten wir das tun?

        Über die Sprecherin
        Judith Simon ist Professorin für Ethik in der Informationstechnologie an der Universität Hamburg. Sie beschäftigt sich mit der Verschränkung ethischer, erkenntnistheoretischer und politischer Fragen im Kontext von Big Data, Künstlicher Intelligenz und Digitalisierung im Allgemeinen. Judith Simon ist Mitglied des Deutschen Ethikrates, sowie verschiedener anderer Gremien für wissenschaftliche Politikberatung. Sie war Mitglied der Datenethikkommission der Bundesregierung (2018-2019).

        Speaker: Prof. Judith Simon (Universität Hamburg)
    • 2:00 PM
      Pause / Break

      Oberes Foyer / Upper Foyer

    • 🇩🇪 Rechenzentrum / Computing Center: Führung 1 / Guided Tour 1 GWDG

      GWDG

      Führung durch das Rechenzentrum der Gesellschaft für Wissenschaftliche Datenverarbeitung Göttingen (GWDG) . Gewinnen Sie einen Einblick in den HPC-/ Technikbereich und erfahren Sie, wo Ihre Daten verarbeitet werden, wenn Sie das KI Servicezentrum für sensible und kritische Infrastrukturen KISSKI nutzen.

      Die angegebenen Zeiten beziehen sich auf Abfahrt und Rückkunft vor dem Tagungsgebäude. Wir bieten einen Shuttleservice mit 2 Kleinbussen. Bitte finden Sie sich kurz vor der Abfahrszeit pünktlich vor dem Gebäude ein. Die Fahrer:innen können nicht warten.

      Die Plätze sind liminiert. Wenn Sie teilnehmen wollen, buchen Sie die kostenpflichtige Führung bei der Anmeldung zur Konferenz.

      Weitere Details hier.

    • 🇩🇪 Keynote Energie. Resilienz in Energiesystemen mit KI - vom Privathaushalt bis zur Netzführung Adam-von-Trott-Saal

      Adam-von-Trott-Saal

      Resilienz in Energiesystemen mit KI - vom Privathaushalt bis zur Netzführung
      (Resilience in energy systems with AI - from private households to grid management)

      Über die Sprecherin
      Prof. Dr.-Ing. Astrid Nieße ist seit 2020 Professorin für Digitalisierte Energiesysteme an der Universität Oldenburg und Mitglied im Bereichsvorstand Energie des OFFIS - Institut für Informatik. Ihr Forschungsschwerpunkt liegt in der Anwendung von informatischen Methoden auf Fragestellungen multimodaler Energiesysteme mit einem Fokus auf Methoden der verteilten künstlichen Intelligenz. Fragestellungen in diesem Kontext betreffen zum Beispiel die agentenbasierte Kontrolle von dezentralen Energieanlagen, selbstorganisierende Systeme, Flexibilitätsmodellierung sowie die sogenannte FAIRification von Forschungsdaten inklusive Forschungssoftware.

      Convener: Prof. Astrid Nieße (Carl von Ossietzky Universität Oldenburg)
      • 7
        🇩🇪 Keynote Energie. Resilienz in Energiesystemen mit KI - vom Privathaushalt bis zur Netzführung

        Resilienz in Energiesystemen mit KI - vom Privathaushalt bis zur Netzführung
        (Resilience in energy systems with AI - from private households to grid management)

        Energiesysteme sind nicht nur zunehmend digitalisiert, sondern stellen auch ein vielversprechendes Anwendungsfeld für Ansätze der Künstlichen Intelligenz dar.
        Im Vortrag werden an konkreten Beispielen vom Privathaushalt bis in die Netzführung unterschiedliche Funktionen in Energiesystemen beleuchtet - welche eignen sich für den Einsatz KI-basierter Verfahren? Und worauf ist zu achten, um die kritische Infrastruktur der Energiesysteme sicher und sichernd zu erhalten?

        Über die Sprecherin
        Prof. Dr.-Ing. Astrid Nieße ist seit 2020 Professorin für Digitalisierte Energiesysteme an der Universität Oldenburg und Mitglied im Bereichsvorstand Energie des OFFIS - Institut für Informatik. Ihr Forschungsschwerpunkt liegt in der Anwendung von informatischen Methoden auf Fragestellungen multimodaler Energiesysteme mit einem Fokus auf Methoden der verteilten künstlichen Intelligenz. Fragestellungen in diesem Kontext sind etwa die agentenbasierten Kontrolle von dezentralen Energieanlagen, agentenbasierte Simulation, Flexibilitätsmodellierung, Methoden der Datenanalyse sowie der Analyse der Resilienz verteilter Energiesysteme.

        Speaker: Prof. Astrid Nieße (Carl von Ossietzky Universität Oldenburg)
    • 🇬🇧 Workshop 1. Explainable Artificial Intelligence (XAI) in Clinical Decision Support Systems Taberna

      Taberna

      Instructors: Jun.-Prof. Dr. Anne-Christin Hauschild, Dr. Zully Ritter (both KISSKI / University Medical Center Göttingen UMG)

      Agenda:

      • Lecture, 30 min
        • Trustworthiness of AI for Clinical Decision Support (CDS) 
        • Introduction to XAI Methods 
        • XAI for CDS 
        • Current Developments on XAI
      • Break, 15 min
      • Hands-on, 120 min:
        • 3 Jupyter labs will be provided for understanding and working the concepts learned in the lecture
        • Jupyter labs:
          (i) ML (setup),
          (ii) ML and exercises on XAI (using an example),
          (iii) complete pipeline for XAI (using an example)
      • Discussion, 15 min

      Max. number of participants: 20

      Conveners: Prof. Anne-Christin Hauschild (Universitätsmedizin Göttingen), Dr Zully Ritter (Universitätsmedizin Göttingen)
    • 🇬🇧 Session 2. Using and Integrating Health Data Emmy-Noether-Saal

      Emmy-Noether-Saal


      Chair: Hendrik Nolte (GWDG), James Bowden (UMG)

      Content/Abstract:
      This session focusses on the integration and processing of health data. We are looking for talks from various related topics which explore the various challenges and some employed solutions for integrating or processing health data.

      Topics of interest include:
      - Integration process of health data with the use of research data management systems (e.g. XNAT, ORTHANC, the Leibniz Data Manager, or a broader data lake)
      - Security and legal aspects (e.g. encryption, anonymization, access control, AAI, pseudonymization, patient consent, enforcing the correct processing of patient consent revocations)
      - How can researchers actually make use of these integrated data sets, (e.g. example use cases when working with XNAT, data stored in the BIDS format, or any other use cases where cross-data source, or cross domain data is processed)
      - Challenges concering processing and scalability

      • 8
        Welcome
      • 9
        FAIR ECG Data processing: From raw data to automated results

        Physiological measurements are permanently acquired in daily clinical practice to aid medical staff in decision making.Electrocardiography (ECG) is one of the most commonly used measurements and represents the physiological activity of the heart. It is widely used in emergency care as well as prevention and in clinical studies. Therefore, it offers large potential for data-driven research, e.g. machine learning for disease prediction in acure care or risk assessment for chronic diseases.

        However, the collected data is often stored in proprietary devices or printed on paper which limits its potential for advanced analysis methods. Moreover, this propreritary nature leads to negative effect w.r.t. open science: Algorithms are oftentimes written once for a specific research project and are dependent on the data formats used, making them non-reusable.

        To overcome these limitations, we aim to develop an interoperable architecture in the framework of the BMBF-funded ACRIBiS project (Advancing Cardiovascular Risk Identification with Structured Clinical Documentation and Biosignal Derived Phenotypes Synthesis). Our goal is to store and exchange ECG data in a findable, accessible, interoperable, and reusable (FAIR) manner. Moreover, we aim for standardized access, selection, filtering, rights management, and user management, enabling scalable use in clinical studies.

        As an initial implementation, the objective is to process signal analysis on ECGs stored in a FAIR manner via the AcuWave Software Suite. We evaluated multiple open source solutions for storing our DICOM ECGs. We initially employed Orthanc, which provides access to the ECGs via its Representational State Transfer - Application Programming Interface. This was utilized to construct a pipeline that can load the ECGs into AcuWave, however, we encountered performance and administrative issues. Consequently, a transition to XNAT is being considered as a potential solution to address these challenges. Within AcuWave, the ECGs are processed in a standardised and modular manner, following a pipeline approach. This architecture will be demonstrated, along with the potential benefits of its automation for research.

        Speaker: Lennart Graf (Institut für Medizinische Informatik, Universitätsmedizin Göttingen)
      • 10
        Towards Integrated Data Management and Automated Analysis of Multi-Omics Datasets in Engineered Human Myocardium

        Engineered human myocardium (EHMs) provides unique opportunity for modeling cardiac development, disease, and drug testing. EHMs, derived from stem cells and cultured over extended periods, closely mimic the structure and function of matured cardiac tissue. This cardiac 3D model offers an innovative platform for exploring cardiac biology and developing novel therapeutic strategies.
        A critical aspect of our research is the integration and automated analysis of multi-omics datasets to elucidate the complex biological processes involved in EHM development, maturation, and response to pharmacological agents. Over a one-year culture period, EHMs are subjected to extensive multi-omics profiling, including proteomics, transcriptomics, and protein turnover studies. High-resolution imaging with SarcAsM, an AI tool, provides detailed structural analysis of sarcomeres, the fundamental contractile units of muscle tissue. Functional performance is regularly assessed using CardioTraceAnalyzer, an AI-driven tool for analyzing time-series data of tissue contractions.
        The integration of these diverse data types presents significant challenges. To address this, we are implementing a Data Lake approach in collaboration with the Gesellschaft für wissenschaftliche Datenverarbeitung Göttingen (GWDG). This strategy ensures that multi-modal data will be accessible, organized, and ready for advanced computational analysis. By centralizing data storage and management, we facilitate seamless data integration, enabling robust cross-disciplinary collaboration and deeper insights into cardiac organoid biology.
        Automated analysis pipelines will be established to process and interpret complex datasets efficiently. These pipelines utilize machine learning algorithms to uncover patterns and correlations within the multi-omics data, driving novel insights into cardiac development and disease mechanisms. Our integrated approach not only advances the understanding of EHM morphogenesis and maturation but also sets a precedent for utilizing large-scale, multi-omics datasets in drug testing and disease modeling. By fostering collaboration with data scientists and leveraging automated analytical techniques, we aim to accelerate the development of innovative treatments for cardiac diseases, ultimately contributing to improved patient outcomes.

        Speaker: Gesine Marie Dittrich (University Medical Center Goettingen)
      • 11
        Improving Cardiovascular Health through AI-based Analysis of Genetic Risk Factors

        Cardiovascular diseases (CVDs) represent a global health challenge, causing approximately 17.9 million deaths worldwide in 2019, including 158,359 deaths in Germany. Traditional risk stratification methods focusing on age, blood pressure, and cholesterol levels often result in inaccurate risk assessments. The goal of stratification is to enable more precise and individualized diagnoses, prognoses, and therapies. By identifying specific subgroups within a broader patient population, medical research can develop targeted and effective treatment strategies tailored to the unique needs of these subgroups in the context of personalized medicine.
        The landscape of personalized medicine, particularly concerning CVDs, has advanced rapidly due to progress in genomics, data analysis, and artificial intelligence (AI). Incorporating non-traditional risk factors, such as genetic information, into clinical practice has the potential to enhance risk prediction accuracy and treatment stratification, significantly aiding in the prevention of complex diseases. For instance, coronary artery disease (CAD), a prevalent CVD, is influenced by multiple genetic variants (alleles) across the genome. The ApoE gene, which encodes the apolipoprotein E protein, plays a critical role in lipid metabolism and is linked to both neurodegenerative and cardiovascular diseases. The ApoE ε4 allele significantly increases the risk of Alzheimer's and CVDs, including CAD. Early identification of individuals with these genetic markers can lead to proactive cholesterol management, substantially reducing the risk of developing CAD.
        Polygenic Risk Scores (PRS), which aggregate the effects of numerous small genetic variations, have shown promise for personalized risk prediction. PRS can be particularly useful for individuals with a family history or genetic predisposition to CVDs. Although PRS are still in the research phase for CVDs, growing studies demonstrate their utility in risk prediction, sparking interest in their clinical application. However, challenges related to data privacy and access to sensitive data persist. The transfer of genomic data from patients to physicians is heavily regulated by privacy laws such as the GDPR and the EU's AI Act.
        This research proposes a patient-centered approach that ensures the confidentiality of individual genomic data. Genetic risk factors will be analyzed using AI to enhance risk prediction and treatment stratification, supporting personalized cardiovascular medicine. The goal is to implement personalized risk prediction and treatment stratification for CVDs through AI-based analysis of genetic risk factors, embedded in a patient-centered approach that guarantees data confidentiality.
        The envisioned approach involves two modules. The patient module, a client-side software on the patient's smartphone, enables local storage of genomic data and local calculation of CVD risk without transmitting genomic data to cloud providers. Patients can select body organs and send risk analysis results to a chosen physician. The analysis identifies potential disease traits and calculates individual PRS, with results condensed by an AI algorithm. Medical results are not displayed to patients but are interpreted by physicians. The physician module allows physicians to access and visualize results, along with summaries of relevant scientific publications. A language model derived from existing open-source large language models will generate controlled summaries, aiding physicians in personalized diagnoses. This approach paves the way for tailored cardiovascular medicine, where data privacy and ethics play central roles. The resulting decentralized, AI-based platform for personalized genomic health services, exemplified by heart diseases, contrasts with traditional cloud-based methods by empowering patients with data control while enabling personalized medicine based on genomic data, expected to enhance patient trust in personalized medical services.

        Speaker: Prof. Wolfgang Maaß
      • 12
        The Leibniz Data Manager - a semantically-driven data catalog

        Research Data Management encompasses the systematic organization, storage, preservation, publication and re-use of data throughout a research project's lifecycle. It aims to ensure effective handling, maintenance, and accessibility of data, promoting reproducibility, transparency, and integrity in scientific research. The surge in digital research data led to the need for standardized practices which are already widely in use in the health domain.
        Within the KISSKI Service Centre, the Leibniz Data Manager (LDM) will provide a semantically-driven data catalog which ensures data transparency and privacy protection while enabling data exchange within the respective research groups and scientific communities. The data catalog is developed at TIB – Leibniz Centre of Science and Technology. To enable shared data from two domains (i.e., health and energy), the LDM resorts to semantic integration using ontologies and standards to transform research, clinical, genomic, and scientific data into semantically rich factual statements. Digital research objects (e.g., data benchmarks) are available following the FAIR data principles. Because strict data access policies regulate health data (e.g., clinical and genomic data), two customized instances of the LDM are deployed representing both, an open data and a closed data system. While the public instance will support the publication and exchange of non-sensitive, anonymized data, the second instance will include a privacy-aware query engine and high security standards provided by partners at GWDG ensuring compliance with data access regulations. For future data use, aka AI-ready data, project partners will define domain-specific ontologies to establish a common understanding of the included medical datasets.

        Speaker: Dr Angelina Kraft (Technische Informationsbibliothek)
      • 13
        GUIDE-IT: Building a medical imaging data sharing platform with XNAT

        Randomized controlled trials (RCTs) are the gold standard for evaluating the efficacy and safety of new medical interventions, including medical imaging technologies that are essential for diagnosis and treatment. The integrity and reliability of RCT results depend heavily on the quality and quantity of data collected. Imaging RCTs have the potential to transform patient care, both in their primary analysis and in subsequent reuse of the imaging and clinical data for secondary analyses. In recent years, there has been a growing recognition of the importance of data sharing. Data sharing plays a critical role in advancing the field of medical imaging by fostering collaboration, transparency, accelerating scientific discoveries and improving patient outcomes. Access to diverse datasets from multiple studies allows researchers to perform large-scale analyses, identify patterns, and uncover new insights that may not be apparent from individual studies. This collaborative approach is driving advances in medical imaging research.

        To this end, in the GUIDE-IT project, we have developed a first prototype of a clinical imaging trials data sharing platform. Using XNAT we provide a platform with fine-grained access control, automated quality control and data processing capabilities. We have evaluated the platform using data from two large coronary imaging RCTs, DISCHARGE and SCOT-HEART.

        Speaker: James Philip Bowden (UKEI)
    • 🇩🇪 Podiumsdiskussion "Künstliche Intelligenz: Turbo für Innovationen im Energiesystem" Adam-von-Trott-Saal

      Adam-von-Trott-Saal


      Über die Mitglieder des Podiums

      Moderation: Reinhard Mackensen (Fraunhofer IEE Kassel)

      Astrid Nieße (Universität Oldenburg)
      Prof. Dr.-Ing. Astrid Nieße ist seit 2020 Professorin für Digitalisierte Energiesysteme an der Universität Oldenburg und Mitglied im Bereichsvorstand Energie des OFFIS - Institut für Informatik. Ihr Forschungsschwerpunkt liegt in der Anwendung von informatischen Methoden auf Fragestellungen multimodaler Energiesysteme mit einem Fokus auf Methoden der verteilten künstlichen Intelligenz. Fragestellungen in diesem Kontext betreffen zum Beispiel die agentenbasierte Kontrolle von dezentralen Energieanlagen, selbstorganisierende Systeme, Flexibilitätsmodellierung sowie die sogenannte FAIRification von Forschungsdaten inklusive Forschungssoftware.

      Lukas Knüsel (BDEW Bundesverband der Energie- und Wasserwirtschaft)
      Lukas Knüsel ist Fachgebietsleiter Digitalisierung beim Bundesverband der Energie- und Wasserwirtschaft (BDEW). Seine Themenschwerpunkte sind die digitale Transformation der Energiewirtschaft, die Digitalisierung der Energiewende sowie aktuelle Entwicklungen in der Daten- und Digitalpolitik.

      Lars Nolting (TenneT)
      Dr. Lars Nolting ist seit 2022 als Projektmanager im Bereich Digital and Flex Development beim Übertragungsnetzbetreiber TenneT tätig. Dort leitet er Projekte zur Digitalisierung im Energiesystem sowie zur Beanreizung von Lastflexibilitäten. Lars Nolting hat im Bereich der Energiesystemanalyse an der RWTH Aachen University promoviert und war als Post-Doc am Lawrence Berkeley National Laboratory in Kalifornien tätig.

      Philipp Richard (dena Deutsche Energie-Agentur)
      Philipp Richard leitet den Bereich Digitale Technologien und Start-up Ökosystem (DTS) in der Deutschen Energie-Agentur (dena). Mit seinem Team verantwortet er die Bearbeitung von Grundsatzfragen für eine nachhaltige Digitalisierung des Energiesystems und möchte das Potenzial von Tech-Start-ups für die Energiewende heben. Schwerpunktsetzung sind dabei u. a. der Einsatz von neuen digitalen Technologien wie Blockchain und künstliche Intelligenz.

      Marc Stanke (DB Systel)
      Dr. Marc Stanke ist Teil des Enterprise Business Management Teams der DB Systel in Frankfurt am Main. Er kommt aus der der Regelung und Steuerung komplexer biotechnologischer Systeme mit Künstliche Intelligenzen und Vorhersagemodellen und hat langjährige Erfahrung in den Bereichen BI, Analytics & Process Mining. Sein Schwerpunkt liegt jetzt auf der Orchestrierung der internen IT und der Etablierung von effektiven, automatisierbaren und wertschaffenden Governance Systemen.

      Conveners: Dr Reinhard Mackensen (Fraunhofer IEE Kassel), Lukas Knüsel (Bundesverband der Energie- und Wasserwirtschaft), Prof. Astrid Nieße (Carl von Ossietzky Universität Oldenburg), Dr Lars Nolting (Tennet), Philipp Richard (Deutsche Energie-Agentur), Dr Marc Stanke (DB Systel)
    • 🇬🇧 Session 1: AutoML: Efficient Development of new AI Applications Hannah-Vogt-Saal

      Hannah-Vogt-Saal


      Chair: Prof. Dr. Marius Lindauer (LUH)

      Description

      Automated Machine Learning (AutoML) is concerned with automating the creation and deployment of machine learning models, enabling non-experts to harness the power of machine learning while also supporting experts by streamlining their workflows.

      This session is divided into three parts. First, we will introduce AutoML, covering its core components and interfaces. We will explore how AutoML automates tasks like model selection, hyperparameter tuning, and neural architecture search, making machine learning more accessible to diverse industries while also improving efficiency for experts by automating repetitive and time-consuming tasks.

      In the second part, we will begin with a brief live demonstration of AutoML tools for users with technical backgrounds. Following this, we will discuss "Green AutoML," an emerging research area focused on the environmental impact of automated machine learning. We will examine research in this field and explore ways to reduce the ecological footprint of machine learning technologies.

      The final part will showcase a no-code AutoML web application, specifically designed to simplify machine learning for non-technical users. Through live demonstrations, we will show how AutoML empowers users in the renewable energy sector to quickly build and deploy models, driving innovation and sustainability in this critical domain.

      • 14
        Tutorial on the basics of AutoML

        Tutorial on the basics of AutoML and how it can support developers in more efficient development of AI applications

      • 15
        Sustainable AutoML
        Speaker: Dr Wever Marcel (Leibniz Universität Hannover)
      • 16
        KI-Lab.EE: Empowering domain experts with No-Code AutoML

        The adoption of specialized AI solutions like predictive maintenance, anomaly detection, and image classification is increasingly important across industries, driving operational efficiency and fostering innovation. However, this requires extensive AI expertise and programming skills. While large organizations can leverage dedicated data science teams, small and medium-sized enterprises (SMEs) face substantial hurdles. No-Code AutoML tools offer a solution, democratizing AI by enabling SMEs and domain experts to implement machine learning applications without requiring extensive AI skills.

        In this lightning talk we will demonstrate how No-Code AI tools can assist non-AI-expert users in creating high-quality AI solutions. The presentation will include an overview of the KI-Lab.EE software, which can be accessed free of charge by SMEs to develop AI-based applications.

        Speaker: Katharina Strecker (Zentrum für Sonnenenergie- und Wasserstoff-Forschung Baden-Württemberg (ZSW))
    • 🇩🇪 Rechenzentrum / Computing Center: Führung 3 / Guided Tour 3 GWDG

      GWDG

      Führung durch das Rechenzentrum der Gesellschaft für Wissenschaftliche Datenverarbeitung Göttingen (GWDG) . Gewinnen Sie einen Einblick in den HPC-/ Technikbereich und erfahren Sie, wo Ihre Daten verarbeitet werden, wenn Sie das KI Servicezentrum für sensible und kritische Infrastrukturen KISSKI nutzen.

      Die angegebenen Zeiten beziehen sich auf Abfahrt und Rückkunft vor dem Tagungsgebäude. Wir bieten einen Shuttleservice mit 2 Kleinbussen. Bitte finden Sie sich kurz vor der Abfahrszeit pünktlich vor dem Gebäude ein. Die Fahrer:innen können nicht warten.

      Die Plätze sind liminiert. Wenn Sie teilnehmen wollen, buchen Sie die kostenpflichtige Führung bei der Anmeldung zur Konferenz.

      Weitere Details hier.

    • 4:50 PM
      Getränke / Drinks

      unteres Foyer

    • 🇬🇧 🇩🇪 Postersession Hannah-Vogt-Saal

      Hannah-Vogt-Saal

      • 17
        Altered EEG network profile after a first unprovoked seizure: an application of a clinical routine big-data sample

        Diagnosing epilepsy after a first unprovoked seizure, especially without visible lesions and with a normal rEEG, is challenging. Understanding EEG network changes requires data collected close to the acute event and follow-up information, necessitating a large data source. The UMG database, with over 34,000 routine EEGs, is invaluable for this purpose. In our study, it took a month to select 42 rEEGs based on specific criteria. Our results indicate that increased connectivity and power in rEEG could serve as biomarkers to predict which patients will develop epilepsy after a first seizure, potentially greatly improving quality of life and reducing treatment costs. This highlights the need for analyzing large datasets to better understand diseases. AI has the potential to significantly reduce data selection time and improve case detection, enabling faster and more efficient analysis.

        Speaker: Marysol Segovia-Oropeza (Clinic of Neurology, University Medical Center, Georg-August University, Göttingen)
      • 18
        Context-Adaptive Patient Data Presentation in CAIS.ME

        Every day, hospital professionals face the challenge of navigating through extensive patient information that is also often not available at the point of care. The complexity and volume of data, along with time pressure, can easily lead to critical information being overlooked. With the digitization of patient records, an adaptive information provision system such as CAIS.ME (Context-Aware Information System for Medical Environments) could support daily clinical work by displaying relevant information at the right time and place on lightweight AR smart glasses, thereby reducing the cognitive load placed on medical staff.

        Given the importance of personalization and good usability in CAIS.ME, one of the major research areas of the project is AI-based customization of the provided data. The presentation style, content, and level of information detail of interest to the user may depend on many factors, including vital signs, medical history, laboratory values, age, other medical conditions, as well as user preferences and context.

        The foundation for the envisioned multi-step frame adaptation process was laid by Tom-Maurice Schreiber in his master's thesis that focused on facilitating decision-making processes, such as finding an optimal treatment, and early detection of alarming changes in a patient's health. The main part of the proposed solution is the knowledge graph that is created on the basis of a small set of clinical guidelines and captures relationships between diseases (e.g., acute appendicitis) and observations (e.g., high white blood cell level), while also taking into account various patient characteristics (e.g., age). Realized in Neo4J as a property graph, it is queried to determine the prioritization of information that is subsequently utilized in the generation of smart glasses frames.

        Speakers: Lidia Roszko, Mr Tom-Maurice Schreiber (TU Dresden)
      • 19
        Enhancing Diagnostic Accuracy in Medicine Through AI

        Artificial Intelligence (AI) has seen a significant surge in popularity, particularly in its application to medicine. The exponential increase in examinations by 3D imaging devices such as CT and MRI has resulted in a massive volume of image data, necessitating the use of AI. Typically, doctors interpret these scans in a time-consuming process often limited by subjectivity, image complexity, inter-observer variability, and fatigue.
        In collaboration with AGEL Radiologia s.r.o. in St. Cyril and Methodius Hospital in Bratislava, we are developing methods for diagnosing various diseases from CT and MR images, focusing on detecting leukoencephalopathy and meningiomas, as well as prostate cancer and kidney stones. Using a dataset of approximately 1,200 patients with axial brain CT scans, we trained convolutional neural networks (CNNs) for binary leukoencephalopathy classification and achieved a classification accuracy of 98.5%. We also achieved 88% accuracy for prostate segmentation and 68% accuracy for prostate cancer classification.
        We emphasize the importance of simplifying dataset creation by healthcare professionals, ensuring that doctors only perform tasks they would typically do during standard diagnoses. We do not require any additional work from doctors to create the labeling for the dataset; instead, we record the data during their routine work. This approach introduces technical challenges in preprocessing, such as converting measured locations into bounding boxes and segmentation based on density differences, which require meticulous technical handling.
        Our research also addresses the critical issue of transparency and interpretability in AI models. To gain insights into model decision-making, we implemented Grad-CAM heatmaps, which highlight the focus areas of the models on the scans. We plan to incorporate other explainable neural network components, such as feature importance and attention mechanisms, to help medical professionals understand why AI systems make specific decisions. This approach increases trust and facilitates the adoption of AI in clinical practice.
        Future research plans include the use of semi-supervised learning methods, which are particularly important when working with medical images due to the scarcity of annotated data. Our plans also include continuous self-learning techniques for AI models, allowing them to adapt and improve as new data becomes available. This approach ensures that the AI system remains up-to-date and maintains high accuracy, which is crucial in healthcare.

        Speaker: Zuzana Cernekova (Comenius University Bratislava)
      • 20
        Flächendeckende Abgrenzung von kleinräumigen Eignungsgebieten für die Wärmeplanung FLAKE

        As part of the project Area-wide delineation of small-scale suitability areas for heat planning (FLAKE), suitability areas for heat supply with heating networks are to be identified. The pro-ject is being carried out in cooperation with GSG Oldenburg. Various geo-AI algorithms will be used to find the best possible area allocations. At the beginning of the project, the data from the housing industry will be merged with other data from municipal heat planning and possible correlations will be investigated. The first algorithm is to find the next most suitable buildings for a heating network along the road network from one building. With the help of the parcels of the buildings, a ring is formed and an area of interest is created. The second algorithm will be based on clustering and create polygons that are as homogeneous as possible and that are well separated from the neighboring polygons. It may be possible to choose between different parameters relevant for heat conduction planning. Following the modeling, the applicability of the algorithms should be evaluated by users without computer science expertise.

        Speaker: Mareike Fincken (Jade Hochschule)
      • 21
        HighREST – a prototype towards a high resolution meteorological weather API with Vision Transformers

        Accurate weather forecasts are core ingredients for many applications.

        Farmers rely on weather forecasts to plan planting, irrigation, and harvesting. Airlines use weather data to plan flight routes, ensure passenger safety, and mini-mize delays. Utility companies use forecasts to anticipate demand changes due to temperature variations and to manage resources efficiently.

        Current state of the art weather models such as ICON from the German weather service have nested regional models such as ICON-EU or ICON-D2 with resolutions from roughly ~6.5 km up to ~2.2 km resolution. The first model provides forecasts for up to 5 days, whereas the second model offers predictions only up to 48 hours in advance. While the second model benefits from higher resolution, it is limited by a shorter forecasting horizon.

        This study builds upon existing weather models by attempting to integrate their strengths, aiming to achieve higher resolution forecasts across both temporal and spatial dimensions. We employed a top-performing Vision Transformer model, train-ing it with one year of ICON data focused on variables such as 2-meter Temperature and 10-meter Wind Speeds using an H100 GPU. Preliminary results are encouraging, demonstrating an improvement over actual measurements.

        With our approach we are aiming to provide a KISSKI service that runs the inference and gives users access to high-resolution forecasts of the two variables for up to 5 days in advance via an API interface.

        Speaker: Gerrit Hein (Fraunhofer IEE)
      • 22
        IntelliLung: AI based decision support system for safe mechanical ventilation in intensive care units

        Mechanical ventilation (MV) is a life-saving therapy used in the intensive care unit (ICU). However, improper settings can lead to lung injury and organ damage. Determining the ideal ventilation settings is challenging due to the large number of variables involved, making it difficult to provide clear guidelines.
        The IntelliLung project addresses this issue with a reinforcement learning (RL) algorithm that was trained on 19 groups of variables, such as blood gas analysis, circulatory function, demographics, and gas exchange, to recommend optimal settings for up to 10 key ventilator parameters. The goal was to maximize ventilator free days during a patient treatment while maintaining safe vital signs. Factored action spaces were used to overcome the challenge of the large number of discrete action combinations. The algorithm has been trained on diverse datasets, including MIMIC IV, eICU, HiRID, and TUD with the cohort consisting of adults on invasive MV. A use case specific preprocessing pipeline has been developed to convert raw medical measurements into a usable state vector using techniques like data imputation, encoding, and measures of central tendency. In safety critical evaluation such as ours, online evaluations are not feasible. Policy evaluation was done offline using Fitted Q-Evaluation (also with factored action space). This evaluation showed that the algorithm is able to achieve higher returns than clinicians across all datasets. In addition domain experts validated the policy action distributions from a medical perspective.

        Speakers: Jason Li (InfAI), Muhammad Hamza Yousuf (InfAI)
      • 23
        KPI Extraction from Maintenance Work Orders—A Comparison of Expert Labeling, Text Classification and AI-Assisted Tagging for Computing Failure Rates of Wind Turbines

        Maintenance work orders are commonly used to document information about wind turbine operation and maintenance. This includes details about proactive and reactive wind turbine downtimes, such as preventative and corrective maintenance. However, the information contained in maintenance work orders is often unstructured and difficult to analyze, presenting challenges for decision-makers wishing to use it for optimizing operation and maintenance. To address this issue, this work compares three different approaches to calculating reliability key performance indicators from maintenance work orders. The first approach involves manual labeling of the maintenance work orders by domain experts, using the schema defined in an industrial guideline to assign the label accordingly. The second approach involves the development of a model that automatically labels the maintenance work orders using text classification methods. Through this method, we are able to achieve macro average and weighted average 𝐹1-scores of 0.75 and 0.85 respectively. The third technique uses an AI-assisted tagging tool to tag and structure the raw maintenance information, together with a novel rule-based approach for extracting relevant maintenance work orders for failure rate calculation. In our experiments, the AI-assisted tool leads to an 88% drop in tagging time in comparison to the other two approaches, while expert labeling and text classification are more accurate in KPI extraction. Overall, our findings make extracting maintenance information from maintenance work orders more efficient, enable the assessment of reliability key performance indicators, and therefore support the optimization of wind turbine operation and maintenance.

        Speaker: Bastian Schäfermeier (GNOI)
      • 24
        mlguide – Methodological Guidance for Applied Machine Learning

        Finding the optimal workflow to clean and prepare data, train a machine learning (ML) model, select between different models, and evaluate the final model properly is a complex task. Novice users can easily be overwhelmed by the variety of available methods for the individual workflow steps like data splitting, metric selection, and model evaluation. Many tools have been created to support specific workflow steps (e.g. hyperparameter tuning), often by automation. However, the current landscape lacks a comprehensive tool that guides researchers with their individual research problems through all ML workflow steps efficiently [1].

        Our new software mlguide aims to fill this gap by providing a user-friendly, interactive platform that facilitates the selection of suitable methods at each stage of the machine learning process. Users can define input characteristics of their dataset (e.g., sample size and task type) and their research question (e.g., whether they want to compare models, evaluate a selected model, or both). Based on this information, mlguide recommends suitable methods for each workflow step, e.g. whether a simple train-test-validation split or nested cross-validation is more suitable for the concrete ML task. These recommendations are backed up by a growing database of evidence from scientific literature.

        mlguide is developed in the context of KI-FDZ, a project that investigates the AI readiness of the German Health Data Lab (HDL; in German Forschungsdatenzentrum Gesundheit, FDZ). Specifically, mlguide will be a central component of the so-called “AI Sandbox”, a user-friendly AI toolbox that aims to allow testing specific user scenarios in a secure processing environment with data protection-compliant data sets.

        The mlguide toolkit consists of three modules. 1) The guidance engine, which is part of the R package mlguide, handles user requests and derives suitable methods. 2) The knowledge base, mlguide.core, stores evidence from scientific literature in a structured format. 3) The graphical user interface, the R Shiny application mlguide.app, allows users to specify their research question and select methods for their research problem. For now, mlguide can handle supervised regression and classification tasks on tabular data. Due to its modular design, future extensions to other task and data types are easily possible.

        Currently, a small team of ML experts is reviewing scientific publications and extracting results in a structured format to fill our knowledge base mlguide.core. As this is a labor-intensive process, we performed initial experiments on using large language models (LLMs) to support the evidence extraction process. Our early results in this regard are promising, but we expect human validation of the LLM output to remain necessary to ensure an accurate evidence extraction. In addition to the current capabilities and architecture of mlguide, we present the status of our LLM experiments and share our vision for an efficient, LLM-enhanced evidence extraction process.

        References
        [1] Detjen, H. et al.; Designing Machine Learning Workflows and Experiments with Ease: A Scoping Review of Interactive Tools. Under preparation. 2024.

        Speaker: Max Westphal (Fraunhofer MEVIS)
      • 25
        Optimizing Renewable Energy Forecasts: An MLOps Conceptual Approach for Scalability

        As the expansion of renewable energies grows, the need for accurate energy forecasts becomes crucial due to the dependency on volatile energy sources. Traditional forecasting systems, which utilize weather data and historical generation data, are challenged by the unique behaviors of individual power plants, lack of data and changing conditions (Yan et al. 2022). To address these challenges, we propose a highly scalable energy prediction system based on MLOps principles to ensure efficient model updates while adhering to appropriate quality criteria. This aims to improve the accuracy and efficiency of renewable energy forecasts, supporting a reliable energy supply in the dynamic energy sector.

        The primary purpose of MLOps is to efficiently facilitate the deployment of ML models into production by eliminating bottlenecks in Development and Operations and automating the workflows (Subramanya et al. 2022). Building on these principles, we aim to develop a service mesh comprising multiple interacting microservices, which include a training service for both experimental and production environments, a forecasting service for operational forecasts, a centralized feature store that houses all necessary data including training, test, operational data, and master data, a model store for archiving models with their parameters and metrics, and a monitoring service to ensure prediction quality and service reliability.

        The main challenge we address is the large number of assets, each requiring individual model predictions, necessitating a highly flexible and scalable ML pipeline. We employ dual training modes for initial and continuous model retraining within the same productive Kubernetes cluster used for inference. Furthermore, a separate model store for logging and tracking supports access at any stage of the ML pipeline (Alla and Adari 2021), complemented by a central feature store that
        enhances flexibility and adaptability by utilizing consistent data interfaces (Dowling 2023). For a high scalability of our inference service, we use Horizontal Pod Autoscaling provided by Kubernetes that automatically updates a workload resource to match request demand (The
        Kubernetes Authors 2024).

        In summary, conventional static methods do not fulfill the growing requirements of increasingly frequent dynamic changes in modern energy systems. We therefore aim to use MLOps concepts to provide scalable services for the continuous improvement of forecasts under changing conditions.

        Literature:
        Alla, Sridhar; Adari, Suman Kalyan (2021): Introduction to MLFlow. In Sridhar Alla, Suman Kalyan Adari (Eds.): Beginning MLOps with MLFlow. Berkeley, CA: Apress, pp. 125–227.

        Dowling, John (2023): What is a Feature Store for Machine Learning? Available online at https://www.featurestore.org/what-is-a-feature-store, updated on 8/11/2023, checked on 5/17/2024.

        Subramanya, Rakshith; Sierla, Seppo; Vyatkin, Valeriy (2022): From DevOps to MLOps: Overview and Application to Electricity Market Forecasting. In Applied Sciences 12 (19), p. 9851.DOI: 10.3390/app12199851.

        The Kubernetes Authors (2024): Horizontal Pod Autoscaling. Available online at https://kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/, updated on 2/18/2024, checked on 5/17/2024.

        Yan, Jie; Möhrlen, Corinna; Göçmen, Tuhfe; Kelly, Mark; Wessel, Arne; Giebel, Gregor (2022): Uncovering wind power forecasting uncertainty sources and their propagation through the whole modelling chain. In Renewable and Sustainable Energy Reviews 165, p. 112519. DOI: 10.1016/j.rser.2022.112519.

        Speaker: Johannes Maximilian Franz Schütz (IEE)
      • 26
        Poster presentation: AI-supported health communication in Plain Language

        According to recent studies in the field of Public Health, over half of the German population indicates having difficulties finding, comprehending, appraising and using health-related information (Schaeffer et al. 2021). The German National Action Plan Health Literacy lists Plain Language as a communicative tool to make health information more accessible to large population groups (Schaeffer et al. 2018). There is a high demand for comprehensibility-enhanced health communication, but few translation resources (Maaß 2020).
        To meet the demand for interlingual translations, meaning the translation between natural languages, AI-based translation tools have long been adopted into the interlingual translation process. In intralingual translation, i. e. the translation from one language variety to another, translation tools are still missing. AI-based intralingual machine translation represents a promising tool to support the production of comprehensible texts in the field of health communication (Deilen et al. 2024a, 2024b), but there are numerous research desiderata regarding its implementation, quality of outputs, and post-editing processes.
        The project KI-gestützte Gesundheitskommunikation (KI-GesKom, AI-supported health communication) is a cooperation between the Research Centre for Easy Language, the German health magazine Apotheken Umschau and SUMM AI. Using the AI-based machine translation tool SUMM AI, we compare machine generated Plain Language translations with their source texts and with professionally translated Plain Language texts. We propose a poster to present the research design and results on the levels 1) correctness, 2) readability, and 3) syntactical complexity (Deilen et al. 2023; Deilen et al. 2024a). In terms of correctness, AI-translated texts need (often extensive) post-editing. The analyses reveal not only spelling errors, but also content-related mistakes like incorrect statements or explanations (Deilen et al. 2024b). AI-generated texts have a higher readability than professional translations, but are more syntactically complex (Deilen et al. 2024a). We conclude that AI-based translation tools are useful for the intralingual translation process. However, the translation tools need adequate training data, and the AI-generated texts need to be professionally post-edited.

        Literature
        Deilen, Silvana; Hernández Garrido, Sergio; Lapshinova-Koltunski, Ekaterina; Maaß, Christiane (2023): Using ChatGPT as a CAT tool in Easy Language translation. In: Štajner, Sanja; Saggio, Horacio; Shardlow, Matthew & Alva-Manchego, Fernando (Hrsg.): Proceedings of the Second Workshop on Text Simplification, Accessibility and Readability (TSAR). Shoumen, Bulgaria: INCOMA Ltd., 1–10.
        Deilen, Silvana; Lapshinova-Koltunski, Ekaterina; Hernández Garrido, Sergio; Maaß, Christiane; Hörner, Julian; Theel, Vanessa; Ziemer, Sopie (2024a): Evaluation of intralingual machine translation for health communication. In: Song, Xingyi; Gow-Smith, Edward; Scarton, Carolina; Cabarrão, Vera; Chatzitheodorou, Konstantinos; Cadwell, Patrick; Lapshinova-Koltunski, Ekaterina; Bawden, Rachel; Sánchez-Cartagena, Víctor M.; Haddow, Barry; Kanojia, Diptesh; Nurminen, Mary; Moniz, Helena; Forcada, Mikel & Oakley, Chris (Hrsg.): Proceedings of the 25th Annual Conference of the European Association for Machine Translation: Volume 1: Research And Implementations & Case Studies. Sheffield: European Association for Machine Translation (EAMT), 467–477.
        Deilen, Silvana; Lapshinova-Koltunski, Ekaterina; Hernández Garrido, Sergio; Maaß, Christiane; Hörner, Julian; Theel, Vanessa; Ziemer, Sopie (2024b): Towards AI-supported Health Communication in Plain Language: Evaluating Intralingual Machine Translation of Medical Texts. In: Demner-Fushman, Dina; Ananiadou, Sophia; Thompson, Paul & Ondov, Brian (Hrsg.): Proceedings of the First Workshop on Patient-Oriented Language Processing @LREC-COLING-2024 (CL4Health): Workshop Proceedings, 44–53.
        Maaß, Christiane (2020): Easy language – plain language – easy language plus: Balancing comprehensibility and acceptability. Berlin: Frank & Timme.
        Schaeffer, Doris/Berens, Eva-Maria/Gille, Svea/Griese, Lennert/Klinger, Julia/Sombre, Steffen de/Vogt, Dominique/Hurrelmann, Klaus (2021): Gesundheitskompetenz der Bevölkerung in Deutschland vor und während der Corona Pandemie: Ergebnisse des HLS-GER 2: Universität Bielefeld, Interdisziplinäres Zentrum für Gesundheitskompetenzforschung.
        Schaeffer, Doris/Hurrelmann, Klaus/Bauer, Ulrich/Kolpatzik, Kai (2018): National Action Plan Health Literacy. Promoting health literacy in Germany. Berlin: KomPart.

        Speaker: Sarah Ahrens (Universität Hildesheim)
      • 27
        Power flow forecasts at transmission grid nodes using Graph Neural Networks

        With the increasing integration of renewable energy and evolving power consumption patterns caused by new consumers like electric vehicles and heat pumps, power flows in the electricity grid have become more fluctuating and weather-dependent, challenging grid stability. Accurate power forecasts are essential for grid operators to ensure reliable grid calculations and planning. We present a novel approach combining Multi-Task Learning with a Graph Neural Network (GNN) to predict vertical power flows at trans-formers linking high and extra-high voltage levels. By leveraging an Embedding Multi-Task Learning strategy, our method captures local variations in power flow characteristics. The embedding captures latent node features, enabling weight sharing across all transformers in the node-invariant GNN while distinguishing individual transformer behaviors. The GNN architecture also accounts for dependencies between transformers, recognizing that power flows in an electricity network are interdependent. The proposed method's effectiveness is validated using two real-world datasets from German Transmission System Operators, covering significant portions of the German transmission grid. Results demonstrate that the Multi-Task Graph Neural Network outperforms both standard Neural Networks and standard GNN in power flow prediction, benefiting from the embedding layer. A sign test confirms that our model significantly reduces test RMSE on both datasets compared to benchmark models.

        Speaker: Dominik Beinert (GNOI)
      • 28
        Sentinel-2 Super-Resolution with Real-ESRGAN using Satellite and Aerial Image Pairs and Color Correction Techniques

        Improving the spatial resolution of satellite images offers considerable potential for a wide range of remote sensing applications. This study investigates the use of an Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN) to improve the resolution of Sentinel-2 satellite imagery, using high-resolution digital orthophotos as ground truth for fine-tuning. The Real-ESRGAN model, which was initially trained on synthetic data, is further refined by combining Sentinel-2 and orthophoto image pairs to achieve a 4-fold upscaling. In the fine-tuning process, a color correction strategy is used to minimize differences due to different sensors.
        Although visual inspections show that the fine-tuned Real-ESRGAN images appear sharper and more detailed than those generated by conventional interpolation methods, the structural similarity index (SSIM) shows a lower agreement compared to these methods. This suggests that although the fine-tuned images are visually appealing, they present challenges in objective metric evaluation due to the different sensor characteristics and illumination conditions. Nevertheless, this method offers a promising solution for high-resolution remote sensing applications, especially when visual clarity and detail are paramount.

        Speaker: Maximilian Kleebauer (University of Kassel / Fraunhofer IEE Kassel)
    • 🇬🇧 🇩🇪 Postersession Emmy-Noether-Saal

      Emmy-Noether-Saal

      • 29
        A detailed architecture characterization of Forward-Forward Convolutional Neural Networks Emmy-Noether-Saal

        Emmy-Noether-Saal

        Backpropagation, the standard for training Convolutional Neural Networks (CNNs), is not biologically plausible due to its reliance on forward and backward passes. In December 2022, Professor Geoffrey Hinton, a pioneer in the field, introduced the Forward-Forward (FF) algorithm as a potential alternative. FF avoids backpropagation throughout the entire network by using two forward passes, preventing hidden layers from receiving information from subsequent ones. In 2023, our team was the first to implement this innovative algorithm on CNNs [1].

        Here, we provide a comprehensive characterization of FF-based CNNs, demonstrating their comparable accuracy to standard backpropagation on benchmark datasets. For instance, on the MNIST dataset, our model achieved a test accuracy of 99.16%. We used explainable artificial intelligence tools, such as Class Activation Maps, to show that FF-based CNNs make classifications based on real spatial features. Additionally, we found that each layer contributes unique and valuable information to the final classification.

        Building on Hinton’s findings, which highlighted the slower convergence of FF-trained fully connected networks compared to backpropagation, we introduced novel strategies to accelerate FF convergence in CNNs without sacrificing accuracy. We took advantage of the versatility of the FF algorithm and, by independently tuning the hyperparameters of each hidden layer, we reduced the number of required epochs from 200 to 40, decreasing the training time by more than 50%.
        FF´s advantages, such as its biological plausibility and lower memory requirements, together with our findings and improvements, suggest the potential to establish FF as a promising alternative to backpropagation for image analysis. We aim our work encourages wider adoption of the FF algorithm, leading to further exploration of this promising new paradigm.

        [1] Scodellaro, R., Kulkarni, A., Alves, F., & Schröter, M. (2023). Training Convolutional Neural Networks with the Forward-Forward algorithm. arXiv preprint arXiv:2312.14924.

        Speaker: Lira Yelemessova (Max Planck Institute for Multidisciplinary Sciences (MPINAT), Göttingen)
      • 30
        An introduction to neuromorphic computing and SpiNNaker Emmy-Noether-Saal

        Emmy-Noether-Saal

        As one of the first of its type in Germany, KISSKI has received a test and development board for the SpiNNaker-2 neuromorphic platform, in preparation for a larger system installation in late 2024. This test board will be made available to interested researchers shortly, and the full platform will be offered as a regular KISSKI service to both academic institutions and industry partners.

        Neuromorphic hardware such as SpiNNaker arises as a result of reaching the limits of Moore's law and the ensuing diversification of available hardware architectures, which achieve better performance in specific applications while sacrificing the generalist nature of traditional CPU configurations. Neuromorphic hardware in particular was initially designed to facilitate the simulation of networks of spiking neurons, which attempt to emulate the behavior of real neurons. The hardware design considerations required to perform efficient neuron simulations, namely many distributed but smaller compute cores, tightly coupled in a message passing network, also make neuromorphic architectures suitable to various other applications. Among these can be counted AI/ML (in particular, artificial or deep neural networks, which share some similarities to spiking neurons in behavior and application), constraint and optimization problems, network and graph simulations, real time image and signal processing, robotics, and more. Additionally, the hardware usually exhibits low power consumption when compared to other architectures.

        This poster will present an introduction to the topic of neuromorphic computing, and showcase the SpiNNaker hardware configuration, software stack, as well as some applications. As such it should be of interest to the wider research and industry community attending KonKIS 24.

        Speaker: Dr Martin Leandro Paleico (Gesellschaft für Wissenschaftliche Datenverarbeitung Göttingen (GWDG))
      • 31
        DOSMo-7B: A Large Language Model Trained Exclusively on German text Emmy-Noether-Saal

        Emmy-Noether-Saal

        We introduce DOSMo-7B, an open 7 billion parameter large language model (LLM) trained on 1T tokens of exclusively German text. DOSMo-7B uses the same architecture as Mistral-7B, paired with a custom tokenizer to maximize the encoding efficiency for German text. In contrast to existing approaches, which typically improve the German skills of LLMs with continued pretaining, we perform from scratch pretraining to explore the potential of training LLMs with only German text. In this technical report, we describe our approach to dataset creation, training, and evaluation of DOSMo-7B.

        Speaker: Maximilian Idahl (L3S)
      • 32
        Fostering Scholarly Knowledge Curation with Multimodal AI: Integrating LLMs, VLMs, and Knowledge Graphs for Explainability and Trust Emmy-Noether-Saal

        Emmy-Noether-Saal

        Scholarly knowledge curation faces challenges due to diverse methodologies across scientific fields. Leveraging Large Language Models (LLMs) like GPT-3.5 and visual models, we enhance AI explainability and trustworthiness in knowledge curation. Our approach integrates LLMs and VLMS with the Open Research Knowledge Graph (ORKG) and employs prompt engineering for accurate data extraction from academic literature. This collaborative framework merges neural capabilities with symbolic knowledge graphs and human expertise, addressing practical challenges and promoting transparent, reliable AI applications in scientific research.

        Speaker: Hassan Hussein (LUH)
      • 33
        Hessian.AI Servicezentrum für künstliche Intelligenz / Hessian.AI Service Center for Artificial Intelligence
      • 34
        Heterogeneous networks for time-series prediction and their neuromorphic implementation Emmy-Noether-Saal

        Emmy-Noether-Saal

        Biological neurons exhibit a large degree of heterogeneity. In a recent work, we have shown that heterogeneity in the timescales of rate-based neurons can be exploited for a better input representation in networks, leading to better performance on various tasks comprising nonlinear transformations of time-shifted input. More specifically, we have used a recurrently balanced network driven by multidimensional chaotic input as a dynamic reservoir, and determined the respectively best linear readout approximation for the aforementioned tasks.

        Here, we employ the widely used Brian 2 library to extend our previous studies to networks of spiking neurons. Compared to conventional networks of rate neurons, spiking neural networks have the benefit of being more biologically realistic as well as more energy efficient due their sparse signal transmission. To efficiently train our spiking network, we use so-called surrogate gradient descent methods, which entail, for example, the convolution of the spikes with an exponentially decaying kernel. The biological realism and the larger number of parameters of the spiking model enable us to investigate in more detail the benefits of heterogeneous timescales for predicting, memorizing, and processing chaotic time series in biological and artificial intelligence systems. Next, we exploit the compatibility of our spiking model with neuromorphic hardware systems. Using our recently developed Brian2Lava package, which connects Brian 2 to the neuromorphic computing framework Lava, we implement the spiking model on Intel's neuromorphic chip Loihi 2. This enables us to test the performance of the model on a system that promises to be extremely energy efficient and scalable. Furthermore, we explore the implementation of our model on other neuromorphic systems such as SpiNNaker 2 and memristive devices.

        Speaker: Dr Jannik Luboeinski (University of Göttingen Medical Center)
      • 35
        Inferring parameters of differential equations via Bayesian modelling Emmy-Noether-Saal

        Emmy-Noether-Saal

        We present a framework using automatic differentiation with JAX to estimate the parameters of a dynamical system within a Bayesian framework and showcase two examples. First, we estimate the time-dependent reproduction number underlying COVID-19 cases in the UK. Second, we infer the most probable times of decision-making in a cognitive experiment. Our framework thereby provides a systematic approach to quantify uncertainty in dynamical systems described by differential equations.

        Speaker: Johannes Zierenberg (MPIDS)
      • 36
        KI-Servicezentrum Berlin-Brandenburg / AI Service Center Berlin-Brandenburg Emmy-Noether-Saal

        Emmy-Noether-Saal

      • 37
        KISSKI: KI-Servicezentrum für sensible und kritische Infrastrukturen / KISSKI: AI-Servicecenter for Critical and Sensitive Infrastructure Emmy-Noether-Saal

        Emmy-Noether-Saal

        The KISSKI project is dedicated to developing a robust AI service hub for critical and sensitive infrastructures, with a focus on the healthcare and energy sectors. Given the societal importance of these sectors, the initiative aims to use AI techniques to address challenges and stimulate innovation in Germany's increasingly complex, application-focused AI research environment, which is inundated with data from a wide range of sensors and actuators.

        While the KISSKI Service and Expertise Center primarily focuses on the healthcare and energy industries, it welcomes questions from other sectors and disciplines. In the energy industry, AI technologies help manage renewable energy resources, integrate sectors, handle consumer administration, and automate grid operations. The increasing number of geographically scattered generators and consumers makes human oversight and control impractical, necessitating the use of smart, data-driven AI-enabled methods.

        The participating organization is an important non-profit institution that provides the necessary support and infrastructure for the effective application of AI approaches in critical infrastructures. This firm offers a wide range of IT services, including collaborative tools, data storage, and scientific computing, tailoring its products to the changing needs of researchers and contributing to the overall success of the KISSKI program.

        The KISSKI program is differentiated by its new approach to leveraging AI for societal benefit, particularly in critical areas like as healthcare and energy.

      • 38
        Neuromorphic computing systems – a comparative approach Emmy-Noether-Saal

        Emmy-Noether-Saal

        Neuromorphic computing systems mimic the structural architecture and computational strategies of the human brain, with the aim to boost the performance of artificial intelligence applications as well as simulations of biological brain dynamics. To reach this goal, several technologies exist, following different approaches.

        Loihi 2, developed by Intel, leverages hardware implementation of spiking neural networks to achieve highly efficient and parallel processing capabilities and represents one of the most modern neuromorphic chips. SpiNNaker 2, a neuromorphic system very recently released by the University of Manchester and TU Dresden, focuses on offering high model flexibility while at the same time enabling highly efficient event-based communication. Both Loihi 2 and SpiNNaker 2 promise to offer tremendous scalability. On the other hand, there are memristive devices, which promise the creation of dense, energy-efficient circuitry due to their ability to emulate synaptic functionalities and retain memory without power supply. Furthermore, a novel approach is given by polymeric dendritic devices, which utilize organic materials to mimic the complex, tree-like structures of biological neurons, promising to offer flexible and potentially more bio-realistic computing systems. Last but not least, BrainScaleS, a mixed analog-digital neuromorphic platform, enables to tremendously accelerate brain-inspired computation by leveraging the real-time dynamics of electronic circuits.

        Here, we present important systems and methodologies used in the field of neuromorphic computing and highlight their advantages and drawbacks. Our work entails comparative analyses between the aforementioned systems (Loihi 2, SpiNNaker 2, memristive devices, polymeric dendritic devices, and BrainScaleS), which shall elucidate the strengths and challenges of each system. By this, we aim to identify how these systems can contribute to advancing our understanding of cognitive processes as well as to technological applications.

        Speaker: Prof. Christian Tetzlaff (University of Göttingen Medical Center)
      • 39
        semares - semantic data integration solutions for AI model development and application Emmy-Noether-Saal

        Emmy-Noether-Saal

        Semares is a data integration platform solution developed by Genevention GmbH. Semares facilitates integration, analysis and exploration of life science data and provides programmatic as well as push-button access for bioinformatics and AI applications.

        Speaker: Thomas Lingner (Genevention GmbH)
      • 40
        Unlocking LLMs: Forgiving Errors, Ethical Considerations, and BIAS Challenges in Domain Adaptation Emmy-Noether-Saal

        Emmy-Noether-Saal

        Large language models (LLMs) have shown promising capabilities in several domains, but “Inherent Bias,” “Data Privacy & Confidentiality,” “Hallucinations,” “Stochastic Parrot,” and “Inadequate Evaluations” limit the LLM’s reliability for direct and unsupervised use. These challenges are exacerbated in complex, sensitive, low-resource domains with scarce large-scale, high-quality datasets. Therefore, this work introduces in-context synthetic data generation through LLMs as a potential technique to mitigate the Bias in machine learning (ML) pipelines and scarce data challenges by undertaking two real-world use cases in the healthcare domain (Mental health) and Agri-Food Sapce (Food Hazard Identification). In this context, this work presents expert-annotated Food-Hazard and Conversational Therapy (MI dialogues) datasets developed through the joint effort of LLM and humans in the loop. The data generation process incorporates accurately engineered prompts through cues and tailored information to ensure high-quality dialogue generation, taking into account contextual relevance and false semantic change. Both datasets are comprehensively evaluated through binary and multiclass classification problems that uncover the ability of LLMs to reason and understand domain intricacies. Our empirical findings demonstrate that the data generated through this rigorous quality control process is both plausible and substantially beneficial in enabling ML techniques to address the targeted biases, thereby supporting the use of LLMs for supervised, task-specific applications in sensitive domains like mental health. Our contributions not only provide the MI community with comprehensive datasets but also valuable insights for using LLMs in plausible text generation in multiple low-resource domains.

        Speaker: Vivek Kumar (Universität der Bundeswehr München)
      • 41
        Using LLMs to evaluate free text answers Emmy-Noether-Saal

        Emmy-Noether-Saal

        With the help of experts from Forschungszentrum Jülich, the u-form publishing house aims to develop an AI tutor for exam preparation. As part of a joint research project with the AI service centre WestAI, various LLMs are being tested and integrated into the publisher’s training programmes.

        The research collaboration is about evaluating free text answers to exam questions with the help of open-source LLMs. On the poster we will present the goals, challenges and results of this project.

        Speaker: Fritz Niesel (WestAI - FZ Jülich)
      • 42
        WestAI-Servicezentrum / KI Service Center WestAI Emmy-Noether-Saal

        Emmy-Noether-Saal

    • 🇩🇪 Rechenzentrum / Computing Center: Führung 3 / Guided Tour 3 GWDG

      GWDG

      Führung durch das Rechenzentrum der Gesellschaft für Wissenschaftliche Datenverarbeitung Göttingen (GWDG) . Gewinnen Sie einen Einblick in den HPC-/ Technikbereich und erfahren Sie, wo Ihre Daten verarbeitet werden, wenn Sie das KI Servicezentrum für sensible und kritische Infrastrukturen KISSKI nutzen.

      Die angegebenen Zeiten beziehen sich auf Abfahrt und Rückkunft vor dem Tagungsgebäude. Wir bieten einen Shuttleservice mit 2 Kleinbussen. Bitte finden Sie sich kurz vor der Abfahrszeit pünktlich vor dem Gebäude ein. Die Fahrer:innen können nicht warten.

      Die Plätze sind liminiert. Wenn Sie teilnehmen wollen, buchen Sie die kostenpflichtige Führung bei der Anmeldung zur Konferenz.

      Weitere Details hier.

    • 6:20 PM
      Stehempfang, Get-Together / Stand-up Reception, Get-Together

      Oberes Foyer / upper foyer, Adam-von-Trott-Saal

  • Thursday, September 19
    • 🇬🇧 Workshop 2. Generative Artificial Intelligence in the Energy Sector Taberna

      Taberna



      Instuctors: André Baier, Dominik Beinert, Dr. Bastian Schäfermeier (all KISSKI / Frauenhofer IEE Kassel), Dr. Dajan Mimic (KISSKI / University of Hannover LUH)

      Abstract: The release of Chat GPT has brought generative artificial intelligence to the forefront of public attention. Identifying the most suitable technologies, tools, and solutions in the current overwhelming dynamic of development can be challenging. This workshop aims to offer guidance and provide participants with the opportunity to collaboratively identify and discuss potential use cases of generative artificial intelligence in the energy sector.
      Goal: Interdisciplinary assessment of the potential of generative AI in the energy industry.
      Result: Prioritized list of use cases for generative AI in the energy sector

      Agenda:

      • Input session (1 hour)

        • Introduction to generative AI methodologies (Schäfermeier)
        • Current use cases of generative AI in the energy sector (Baier)
        • Analysis of potential and risks of generative AI in the energy sector (Baier)
      • Interactive session (2 hours)

        • Break out 1: Brainstorming additional use cases
        • Plenary discussion
        • Break out 2: Brainstorming additional use cases
        • Plenary discussion
        • Break out 3: Finalization of selected use cases
        • Plenary discussion
        • Summary

      Target Group: Energy Experts, Generative AI Experts, Generative AI interested

      Max. number of participants: 20

    • 🇩🇪 Begrüßung / Welcome Adam-von-Trott-Saal

      Adam-von-Trott-Saal

      Convener: Prof. Julian Kunkel (GWDG / Georg-August-Universität Göttingen)
    • 🇩🇪 Rechenzentrum / Computing Center: Führung 4 / Guided Tour 4 GWDG

      GWDG

      Führung durch das Rechenzentrum der Gesellschaft für Wissenschaftliche Datenverarbeitung Göttingen (GWDG) . Gewinnen Sie einen Einblick in den HPC-/ Technikbereich und erfahren Sie, wo Ihre Daten verarbeitet werden, wenn Sie das KI Servicezentrum für sensible und kritische Infrastrukturen KISSKI nutzen.

      Die angegebenen Zeiten beziehen sich auf Abfahrt und Rückkunft vor dem Tagungsgebäude. Wir bieten einen Shuttleservice mit 2 Kleinbussen. Bitte finden Sie sich kurz vor der Abfahrszeit pünktlich vor dem Gebäude ein. Die Fahrer:innen können nicht warten.

      Die Plätze sind liminiert. Wenn Sie teilnehmen wollen, buchen Sie die kostenpflichtige Führung bei der Anmeldung zur Konferenz.

      Weitere Details hier.

    • 🇩🇪 Keynote Medizin - Replizierbares maschinelles Lernen in der Medizin Adam-von-Trott-Saal

      Adam-von-Trott-Saal

      Replizierbares maschinelles Lernen in der Medizin
      (Replicable machine learning in medicine)

      Über die Sprecherin
      Anne-Laure Boulesteix erwarb ein Diplom in Ingenieurwissenschaften an der Ecole Centrale Paris, ein Diplom in Mathematik an der Universität Stuttgart und einen Doktortitel in Statistik an der Ludwig-Maximilians-Universität (LMU) München. Nach einer Postdoc-Phase in medizinischer Statistik wechselte sie als Juniorprofessorin und Professorin (2012) an die Medizinische Fakultät der LMU. Sie arbeitet an der Schnittstelle zwischen Biostatistik, maschinellem Lernen und Medizin mit einem besonderen Fokus auf Metawissenschaft und Methodenevaluation. Sie ist Mitglied im Lenkungsausschuss der STRATOS-Initiative, Gründungsmitglied des LMU Open Science Center und Präsidentin der deutschen Region der International Biometric Society.

      Convener: Prof. Anne-Laure Bolesteix (Ludwig-Maximilians-Universität München)
      • 43
        🇩🇪 Keynote Medizin. Replizierbares maschinelles Lernen in der Medizin Adam-von-Trott-Saal

        Adam-von-Trott-Saal

        Replizierbares maschinelles Lernen in der Medizin
        (Replicable machine learning in medicine)

        Über die Sprecherin
        Anne-Laure Boulesteix erwarb ein Diplom in Ingenieurwissenschaften an der Ecole Centrale Paris, ein Diplom in Mathematik an der Universität Stuttgart und einen Doktortitel in Statistik an der Ludwig-Maximilians-Universität (LMU) München. Nach einer Postdoc-Phase in medizinischer Statistik wechselte sie als Juniorprofessorin und Professorin (2012) an die Medizinische Fakultät der LMU. Sie arbeitet an der Schnittstelle zwischen Biostatistik, maschinellem Lernen und Medizin mit einem besonderen Fokus auf Metawissenschaft und Methodenevaluation. Sie ist Mitglied im Lenkungsausschuss der STRATOS-Initiative, Gründungsmitglied des LMU Open Science Center und Präsidentin der deutschen Region der International Biometric Society.

        Speaker: Prof. Anne-Laure Boulesteix (Ludwig-Maximilians-Universität München)
    • 🇩🇪 Podiumsdiskussion "KI-Einsatz im Gesundheitswesen: Aktuelle technische, rechtliche und ethische Entwicklungen (AI Act), Herausforderungen und Möglichkeiten" Adam-von-Trott-Saal

      Adam-von-Trott-Saal


      Über die Mitglieder des Podiums

      Moderation: Dagmar Krefting (Universitätsmedizin Göttingen)
      Prof. Dagmar Krefting ist Direktorin des Instituts für Medizinische Informatik an der Universitätsmedizin Göttingen und hat eine Professur für Medizinische Informatik inne. Ein Schwerpunkt ihrer Arbeit liegt auf der institutionenübergreifenden gemeinsamen Nutzung und Analyse von multidimensionalen Biosignalaufzeichnungen wie EKG und EEG, aber auch von Sensordaten, die über Wearables gesammelt werden. Ein weiterer Schwerpunkt liegt auf der Sicherheit, Reproduzierbarkeit und Zuverlässigkeit der Forschungsumgebungen, einschließlich Deep-Learning-Methoden.

      Malte Schmieding (Bundesgesundheitsministerium)
      Dr. Malte Schmieding ist Referent für Künstliche Intelligenz im Referat 511 Neue Technologien und Datennutzung des Bundesministeriums für Gesundheit. Er wirkte in dieser Rolle unter anderem an der EU KI-Verordnung (AI Act) und dem Gesundheitsdatennutzungsgesetz (GDNG) mit. Zuvor arbeitete als Wissenschaftlicher Mitarbeiter am Institut für Medizinische Informatik der Charité – Universitätsmedizin Berlin und als Arzt in Weiterbildung für Allgemeinmedizin in Bottrop. Nach seinem Medizinstudium in Berlin studierte er Biomedical Informatics an der Harvard Medical School, Boston.

      Udo Schneider (Techniker Krankenkasse)
      Dr. Udo Schneider studierte Volkswirtschaftslehre an der Universität Mannheim. Im Anschluss daran war er Mitarbeiter an der Universität Greifswald und promovierte über das Thema Anreize in der Arzt-Patient-Beziehung, bevor er an die Universität Bayreuth wechselte, wo er sich 2008 habilitierte. Von 2012 bis 2018 war er im WINEG (Wissenschaftliches Institut der TK) im Bereich der Analyse von Routinedaten tätig. Seit Mitte 2018 arbeitet er im Fachbereich Versorgungsmanagement der TK und beschäftigt sich mit Versorgungsforschung auf Basis von GKV-Routinedaten, Versorgungsinnovationen und digitaler Versorgung.

      Larisa Wewetzer (Ottobock)
      Dr. Larisa Wewetzer blickt auf eine 13-jährige Erfahrung in der Medizintechnikindustrie zurück. In ihrer derzeitigen Position bei Ottobock, einem führenden Unternehmen in der Branche, leitet sie die globale Entwicklung für Digital Health Solutions. Ihr Hauptaugenmerk liegt dabei auf der Entwicklung und Einführung eines digitalen Produktangebots, das gezielt auf die Anforderungen von Nutzern von Prothesen und Orthesen ausgerichtet ist. In den vergangenen vier Jahren widmete sie sich zudem intensiv der Erforschung der Faktoren, die die Implementierung von KI-gestützten Systemen im deutschen Gesundheitswesen beeinflussen. Ihre Arbeit verbindet wissenschaftliche Erkenntnisse mit praktischer Anwendung, was ihr ein umfassendes Verständnis für die spezifischen Bedürfnisse in der Medizintechnik ermöglicht.

      Helena Zacharias (Medizinische Hochschule Hannover)
      Prof. Dr. Helena U. Zacharias ist Professorin für “Klinische Datenwissenschaften” am Peter L. Reichertz Institute für Medizinische Informatik an der Medizinischen Hochschule Hannover. Sie studierte Physik an der Universität Regensburg und promovierte in Biologie am Institut für Funktionelle Genomik, Universität Regensburg, mit einem Fokus auf Kernspinresonanz-Spektroskopie-basierten Metabolomics-Studien. In ihrer PostDoc-Zeit arbeitete sie an der Ohio State University und am Helmholtz Zentrum München. Von 2019-2021 leitete sie die Nachwuchsgruppe “Computational biomarker discovery” an der Universitätsmedizin Greifswald, und von 2021-2022 war sie Professorin für “Klinische Metabolomics” in der Klinik für Innere Medizin I/Institut für klinische Molekularbiologie an der Universität Kiel und dem Universitätsklinikum Schleswig-Holstein. Ihre Forschungsschwerpunkte liegen in der verbesserten Prädiktion von adversen Patientenereignissen und der Untersuchung metabolischer Pathomechanismen, die komplexen Krankheiten zugrundeliegen.

      Jana Zschüntzsch (Universitätsmedizin Göttingen)
      Priv.-Doz. Dr. Jana Zschüntzsch ist Oberärztin in der Klinik für Neurologie und Leiterin der Neuromuskulären Ambulanz der Klinik für Neurologie sowie Sprecherin des neuromuskulären Zentrums der UMG. Sie war initiativ an der Konzeption des europaweite, interdisziplinäre EU-Forschungsprojekt „Screen4Care“ beteiligt und arbeiten jetzt aktiv an den Zielen und der Umsetzung des innovativen Forschungsvorhabens mit. „Screen4Care“ hat sich das Ziel gesetzt, den Weg bis zur Diagnosestellung von seltenen Erkrankungen zu verkürzen. In „Screen4Care“ leitet Priv.-Doz. Dr. Jana Zschüntzsch das „digitale“ Arbeitspaket zum Einsatz von Algorithmen, die auf maschinellem Lernen beruhen. Patient:innen im frühen Krankheitsstadium sollen so über elektronische Patientenakten schneller identifiziert werden können, damit ihnen auf ihrer Diagnosereise zeitnah geholfen werden kann.

      Conveners: Prof. Dagmar Krefting (Universitätsmedizin Göttingen), Dr Malte Schmieding (Bundesministerium für Gesundheit), Dr Udo Schneider (Techniker Krankenkasse), Dr Larisa Wewetzer (Ottobock), Prof. Helena Zacharias (Medizinische Hochschule Hannover), Prof. Jana Zschüntzsch (Universitätsmedizin Göttingen)
    • 🇬🇧 Session 3. Large Language Models Hannah-Vogt-Saal

      Hannah-Vogt-Saal


      Chair: Dr. Jennifer D'Souza (TIB Hannover)

      Content/Abstract:
      This session will delve into the intricacies of Large Language Models (LLMs). The session aims address the impact of these AI systems along the topics of interest (see below), exploring their capabilities, implications, and potential applications.

      Topics of interest:
      - Pretraining Techniques for LLMs: Exploring foundational strategies and algorithms in the development of LLMs.
      - Testing and Evaluating LLM Fitness:Methods for assessing LLM performance on well-known tasks and benchmarks.
      - Application of LLMs in Scientific Research: Case studies and examples of LLMs driving discovery and innovation in various scientific fields.
      - Innovative Insights Generation: Strategies for leveraging LLMs to generate novel insights and accelerate research outcomes.
      - Challenges and Solutions in LLM Application: Discussing the practical challenges encountered in applying LLMs to scientific research and potential solutions.

      • 44
        Extracting Biochemical Knowledge Graphs from Academic Literature through Large Language Models

        The growth and evolution of our civilization have been based on the use of biodiversity. Throughout history, human survival has been intricately intertwined with the utilization of diverse plant species, serving as vital sources of nutrition and medicinal remedies. The exploration of natural products is a cornerstone in the quest for innovative pharmaceuticals. A staggering 67% of all globally approved drugs trace their origins back to natural sources, whether as pure compounds, semi-synthetic derivatives, or bioinspired pharmacophores.

        Natural product researchers have been studying the medicinal properties of plants, and significant advances regarding biosynthesis, ecology, and biological properties were achieved in the last century. The biodiversity of tropical and equatorial environments is plentiful. It could offer a particularly rich potential in the search for biologically active compounds to be used as models for product discovery and medicinal chemistry. This extraordinary biodiversity and its rich chemical diversity hold immense potential for the development of bioproducts, including pharmaceuticals, cosmetics, food supplements, and agricultural pesticides.

        In this talk, we will shed light on our ongoing effort to bring order and accessibility to the biochemical information available in a structured and organized manner, automated through Large Language Models, to unlock its full potential. Our final goal is to reduce the time spent on scientific studies and processes involving technological development and research of sustainable products and assist in the development of different fields of science, technological development of bio-friendly and sustainable products with high added value, as well as the establishment of novel public policies.

        Speaker: Edgard Luiz Marx (InfAI)
      • 45
        Towards a Reliable Web of Knowledge

        The advent of Large Language Models (LLMs), most notably ChatGPT, has fascinated researchers and the public alike and LLMs. The main attractor toward LLMs is their capability to interpret prompts formulated in natural language and to respond accordingly, allowing for more organic interactions with LLM-based AI systems and increasing their accessibility especially for less tech-savvy users. LLMs gain these capabilities by being trained on a huge corpus of texts and, in the process, learning about patterns of knowledge encoded in this text corpus.

        However, we now know that the truth is more complex. Namely, ChatGPT was shown to make mistakes and hallucinate if a response is statistically likely but not conforming to actual knowledge. While arguably harmless in settings motivated by curiosity and learning about the technology and its capabilities, such mistakes are prone to causing real-world damage if they occur in professional settings, become harder to identify, and actively influence our communication and decision-making processes.
        Fundamentally, we argue that LLMs should be increasingly seen as what their name implies: AI models for understanding and generating human languages, which provide intuitive human-computer interfaces. Conversely, at least for critical applications, they should not be seen as knowledge models.

        This perspective is motivated by the recent shift toward deploying special-purpose LLMs and finetuning existing LLMs for singular applications. This approach is sensible from the perspective of the application provider: They can seize the language-processing capabilities of advanced LLMs, such as ChatGPT, and provide application-specific knowledge in the form of additional training data.

        However, this development disconnects from the global view of a universally connected Internet. In a sense, the development outlined above is even antithetical to the evolution of the Internet: In the era before search engines, users would have to know relevant websites; the advent of search engines provided them with a central hub for finding relevant information. Given that LLMs are at least partly used to get information, the above-mentioned shift toward a vast landscape of application-specific LLMs bears the potential of partly reverting this convenience if users now have to know which LLM serves their purpose best.

        We argue that, from an end-user perspective, we should strive for establishing an LLM-based Web of Knowledge (WoK), which uses an LLM with versatile interpretation capabilities as a unified interface for querying relevant knowledge pools before formulating an answer. In contrast to implicit knowledge learning or relying on a federation of specialized LLMs, an LLM-based WoK could rather analyze users’ prompts for relevant knowledge repositories and gain the information from there in a structured manner before formulating an answer in natural language.

        On the flip side, the idea of a WoK introduces new challenges as well, which require additional attention, such as well-defined APIs for information retrieval or the combination of knowledge obtained from different sources. In this presentation, we outline the concept of the WoK and point out further research efforts required to steer toward a WoK.

        Speaker: Roman Matzutt (Fraunhofer FIT)
      • 46
        Knowledge Injection Strategies for Large Language Models

        During pre-training a Large Language Model (LLM) learns language features such as syntax, grammar and, to a certain degree, semantics. In this process, the model not only acquires language, but also implicitly acquires knowledge. In this sense knowledge is a byproduct of language acquisition. This characteristic is inherent in the architecture of modern LLMs. Hence, much like language features, knowledge is also learned in a fuzzy manner by these models, leading to difficulties in understanding highly specific concepts and a tendency to generate hallucinations.

        This talk will explore various techniques and strategies for sharpening fuzzy knowledge in LLMs with domain-specific information. Here, our focus is on lightweight methods, that require no further pre-training. We will examine direct text injection strategies for LLM encoders (such as triple injection and K-BERT), integration of additional features using multilayer perceptrons (MLPs), and the use of modular lightweight components with knowledge-based adapters. Furthermore, we will investigate injection techniques for LLM decoders that go beyond simple prompting, such as RAGs, and how these can be enhanced by a multi-agent architecture.

        Speaker: Benjamin Wolff (ZB MED - Information Centre for Life Sciences)
      • 47
        Enhancing Emotional Support Chatbots for Depression Using Reinforcement Learning with Expert Knowledge Integration

        Abstract

        Mental health has become a paramount concern in contemporary society, as an increasing number of individuals are grappling with depression​ (WHO, 2023)​. People increasingly rely on partners, friends, or even chatbots as effective means to prevent and alleviate symptoms of depression​ (Rauws, 2022)​. When observing interpersonal interactions, it's evident that conversations are frequently driven by emotional responses rather than logical reasoning. For example, in reaction to a statement like, “My teacher is so annoying. He insists that I attend his course precisely on time”, a supportive friend might empathize by saying, “That sounds tough. Everyone has urgent matters sometimes. Being 10 minutes late shouldn't matter”. In contrast, a chatbot, designed to respond logically, might suggest, “It seems important to adhere to the schedule as required by your course guidelines”.

        To enhance the effectiveness of an emotional support chatbot specifically designed for supporting individuals with depression, integrating proven therapeutic approaches like Solution-Focused Brief Therapy (SFBT) and Cognitive Behavioral Therapy (CBT) can be beneficial. A practical approach is to involve developing datasets from expert therapeutic sessions to train the chatbot to ensure it learns to emulate effective therapeutic communication. Additionally, integrating therapeutic principles into the advanced layers of transformer models can allow the chatbot to generate responses that are not only contextually appropriate but also therapeutically beneficial ​(Qi Ge, 2023)​. This method enhances the chatbot's capability to provide meaningful support, making it a valuable tool in mental health care. However, these techniques pose the risk of catastrophic forgetting, where the model may lose earlier knowledge as it acquires new information.

        We aim to examine the application of Reinforcement Learning (RL) to embed expert therapeutic knowledge in chatbots, focusing on Solution-Focused Brief Therapy (SFBT) and Cognitive Behavioral Therapy (CBT). By employing Inverse Reinforcement Learning (IRL), we deduce the reward functions implicit in annotated expert behaviors. This allows a pre-trained Large Language Model (LLM) to be fine-tuned through RL. The goal is to enable the chatbot to conduct multi-turn emotional support conversations that more effectively support and enhance the patient's emotional well-being, while mitigating the issue of catastrophic forgetting.

        ​​References

        ​​WHO (2023). Depressive disorder (depression). Retrieved from WHO Official Website: https://www.who.int/news-room/fact-sheets/detail/depression

        ​Qi Ge, L. L. (2023). Designing Philobot: A Chatbot for Mental Health Support with CBT Techniques. Proceedings of 2023 Chinese Intelligent Automation Conference .

        ​Rauws, M. (2022). The Rise of the Mental Health Chatbot. Artificial Intelligence in Medicine, 1609–1618.​​

        Speaker: Lingxiao Kong (Fraunhofer Institute for Applied Information Technology FIT)
      • 48
        Unlocking LLMs: Forgiving Errors, Ethical Considerations, and BAIS Challenges in Domain Adaptation

        Large language models (LLMs) have shown promising capabilities in several domains, but “Inherent Bias,” “Data Privacy & Confidentiality,” “Hallucinations,” “Stochastic Parrot,” and “Inadequate Evaluations” limit the LLM’s reliability for direct and unsupervised use. These challenges are exacerbated in complex, sensitive, low-resource domains with scarce large-scale, high-quality datasets. Therefore, this work introduces in-context synthetic data generation through LLMs as a potential technique to mitigate the Bias in machine learning (ML) pipelines and scarce data challenges by undertaking two real-world use cases in the healthcare domain (Mental health) and Agri-Food Sapce (Food Hazard Identification). In this context, this work presents expert-annotated Food-Hazard and Conversational Therapy (MI dialogues) datasets developed through the joint effort of LLM and humans in the loop. The data generation process incorporates accurately engineered prompts through cues and tailored information to ensure high-quality dialogue generation, taking into account contextual relevance and false semantic change. Both datasets are comprehensively evaluated through binary and multiclass classification problems that uncover the ability of LLMs to reason and understand domain intricacies. Our empirical findings demonstrate that the data generated through this rigorous quality control process is both plausible and substantially beneficial in enabling ML techniques to address the targeted biases, thereby supporting the use of LLMs for supervised, task-specific applications in sensitive domains like mental health. Our contributions not only provide the MI community with comprehensive datasets but also valuable insights for using LLMs in plausible text generation in multiple low-resource domains.

        Speaker: Vivek Kumar (Universität der Bundeswehr München)
      • 49
        Steering Large Language Models (LLMs) using Sparse Autoencoders (SAEs)

        A brief into on how Sparse Autoencoders (SAE) can be leveraged to extract interpretable, monosemantic features from the opaque intermediate activations of LLMs, providing a window into their internal representations. And we hope to initiate discussions on the methodology of training SAEs on LLM activations, the resulting sparse and high-dimensional representations, and how these can be utilized for model steering tasks.
        We’ll examine a case study demonstrating the effectiveness of this approach in changing the level of model “proficiency”. This discussion aims to highlight the potential of SAEs as a scalable, unsupervised method for disentangling LLM behaviors, contributing to the broader goals of AI interpretability and alignment.

        Speaker: Lalith Manjunath (GNOI)
    • 🇬🇧 Session 4. Large AI Models by and for Europe Emmy-Noether-Saal

      Emmy-Noether-Saal


      Chair: Dr. Wolfgang Stille (hessian.AI)

      Content/Abstract:
      This session examines two critical aspects of European AI: sovereignty and safety. First, it highlights the European LLM landscape, featuring an open-source language model initiative and advancements in code language models. Second, it explores approaches to address safety and fairness concerns, as well as model safeguarding techniques.

      Agenda:
      - Welcome: 5 min
      - 4 contributed talks (each 15 min + 5 min discussion)

      • 50
        Safety and Fairness in Large AI Models

        In the burgeoning field of artificial intelligence, ensuring the safety and fairness of larg-scale models is paramount. This talk presents Aurora-M, a state-of-the-art multilingual model designed specifically for non-English, e.g. European, languages. Our focus lies in the safety tuning process of Aurora-M, highlighting innovative methodologies to mitigate biases and enhance the model's ethical integrity. We will delve into the broader implications of safety and fairness in large AI models, discussing best practices and challenges faced in the European context. Join us as we explore how Aurora-M exemplifies the commitment to creating AI that is not only powerful and versatile but also aligned with the values of fairness and safety.

        Speaker: Felix Friedrich (Hessian.AI, TU Darmstadt)
      • 51
        Distilling Graph Structure Knowledge into Code Language Models for Generative Tasks

        Although pre-trained language models (PLMs) on code get significantly better, code is largely treated as sequential.
        By ignoring easily extractable structural rules -- through static analysis -- that programming languages and algorithmic concepts follow, significant potential for improvement is lost.
        Some previous work used abstract syntax trees (ASTs) and their extended versions by extracting paths, flattening or using graph-based auxiliary training, etc., which have been shown to improve either performance or reliability in code generation and program synthesis.
        Most of these methods disrupt the nature of graph modality by applying such adaptations to work with transformer-based sequential models, which are currently state of the art.
        We propose a novel method to work directly with graph representations by using graph neural networks (GNNs) and infusing learned structural information into sequential transformer models.
        In doing so, the learned structural knowledge from GNNs is distilled into PLM to help with generative tasks where the target is a programming language and therefore there are no graphs, as opposed to code summarization or search tasks.

        Speaker: Mert Tiftikci (Hessian.AI, TU Darmstadt)
      • 52
        Llavaguard: VLM-based Safeguards for Vision Dataset Curation and Safety Assessment

        We introduce LlavaGuard, a family of VLM-based safeguard models, offering a versatile framework for evaluating the safety compliance of visual content.
        Specifically, we designed LlavaGuard for dataset annotation and generative model safeguarding.
        To this end, we collected and annotated a high-quality visual dataset incorporating a broad safety taxonomy, which we use to tune VLMs on context-aware safety risks.
        As a key innovation, LlavaGuard's responses contain comprehensive information, including a safety rating, the violated safety categories, and an in-depth rationale.
        Further, our introduced customizable taxonomy categories enable the context-specific alignment of LlavaGuard to various scenarios.
        Our experiments highlight the capabilities of LlavaGuard in complex and real-world applications.
        We provide checkpoints ranging from 7B to 34B parameters demonstrating state-of-the-art performance, with even the smallest models outperforming baselines like GPT-4.
        We make our dataset and model weights publicly available and invite further research to address the diverse needs of communities and contexts.

        Speaker: Lukas Helff (Hessian.AI, TU Darmstadt)
      • 53
        Occiglot - Open Language Models by and for Europe

        Large language models (LLMs) have emerged as transformative tools, revolutionizing various natural language processing tasks. Despite their remarkable potential, the LLM landscape is predominantly shaped by US tech companies, leaving Europe with limited access and influence. This talk will present Occiglot - an ongoing research collective for open-source language models for and by Europe. More specifically, we will explain why open European LLMs are needed and share insights as well as lessons learned, ranging from data collection and curation, model training and evaluation.

        Speaker: Malte Ostendorff (German Research Center for Artificial Intelligence (DFKI), Occiglot)
      • 54
        Hosting of LLMs by KISSKI - Easy Access in Germany
        Speaker: Ali Doosthosseini (GWDG)
    • 🇩🇪 Rechenzentrum / Computing Center: Führung 5 / Guided Tour 5 GWDG

      GWDG

      Führung durch das Rechenzentrum der Gesellschaft für Wissenschaftliche Datenverarbeitung Göttingen (GWDG) . Gewinnen Sie einen Einblick in den HPC-/ Technikbereich und erfahren Sie, wo Ihre Daten verarbeitet werden, wenn Sie das KI Servicezentrum für sensible und kritische Infrastrukturen KISSKI nutzen.

      Die angegebenen Zeiten beziehen sich auf Abfahrt und Rückkunft vor dem Tagungsgebäude. Wir bieten einen Shuttleservice mit 2 Kleinbussen. Bitte finden Sie sich kurz vor der Abfahrszeit pünktlich vor dem Gebäude ein. Die Fahrer:innen können nicht warten.

      Die Plätze sind liminiert. Wenn Sie teilnehmen wollen, buchen Sie die kostenpflichtige Führung bei der Anmeldung zur Konferenz.

      Weitere Details hier.

    • 11:15 AM
      Pause / Break
    • 🇩🇪 Rechenzentrum / Computing Center: Führung 6 / Guided Tour 6 GWDG

      GWDG

      Führung durch das Rechenzentrum der Gesellschaft für Wissenschaftliche Datenverarbeitung Göttingen (GWDG) . Gewinnen Sie einen Einblick in den HPC-/ Technikbereich und erfahren Sie, wo Ihre Daten verarbeitet werden, wenn Sie das KI Servicezentrum für sensible und kritische Infrastrukturen KISSKI nutzen.

      Die angegebenen Zeiten beziehen sich auf Abfahrt und Rückkunft vor dem Tagungsgebäude. Wir bieten einen Shuttleservice mit 2 Kleinbussen. Bitte finden Sie sich kurz vor der Abfahrszeit pünktlich vor dem Gebäude ein. Die Fahrer:innen können nicht warten.

      Die Plätze sind liminiert. Wenn Sie teilnehmen wollen, buchen Sie die kostenpflichtige Führung bei der Anmeldung zur Konferenz.

      Weitere Details hier.

    • 🇬🇧 Session 5. AI-Applications in Clinical Practice: Challenges and Successes Adam-von-Trott-Saal

      Adam-von-Trott-Saal


      Chair: Jun.-Prof. Dr. Anne-Christin Hauschild (UMG), Dr. Nicolas Spicher (UMG), Dr. Zully Ritter (UMG)

      Content / Abstract:
      Nowadays, AI applications in healthcare and in general for clinical practice have proved to be successful in specific tasks like diagnosis prediction, risk estimation (e.g., heart failure risks), alarm and order automation, or cancer differential diagnosis; on the other hand, these achievements are confronted to endurance challenges related mainly to patient security (data privacy and data ownership) and bias additionally to ethical issues. Even without considering the regulation of AI and the coming AI Act, AI solutions to be implemented in clinical practice have been developed in academia and industry. In our session, we will provide insights into already proven solutions and those being tested to be efficiently implemented in clinical setups. We will present and delve into machine-learning approaches using mainly clinical, image, and biosignal data. Concerning solutions to handle data privacy properly, among other patient security issues, we will present in this session how, for example, federated learning is gaining in importance as an appropriate technique, allowing not only to solve the problems related to data sharing (both ethical and technical) but also even improving the performance of AI solutions in the case of not having enough or representative data to find a good performing AI solution.
      Aiming to share and discuss the lessons learned during designing and testing AI solutions for clinical practice, we invite you to participate and be part of it.

      Topics of interest include:
      - Lessons learned from AI projects in clinical practice
      - Federated Learning as a novel modality enabling AI in clinical practice
      - Translation of AI methods for biosignal processing towards clinical practice

      • 55
        Topic Introduction
        Speaker: Prof. Anne-Christin Hauschild (Universitätsmedizin Göttingen)
      • 56
        Keynote: Augment my Doc? Chances and Risks of AI in Medicine

        Artificial Intelligence has a high potential of improving medical patient treatment by augmenting the intelligence of human doctors, supporting them during critical tasks or just making their lives easier by partially taking over time-consuming duties. In this talk, we’re going to imagine how AI support would look like in a perfect state, have a look at current approaches of AI in medicine and elaborate on the challenges and risks to incorporate them in practice.

        Speaker: Michael Dietrich (Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI))
      • 57
        Predicting heart failure recovery by wearables and machine learning

        Heart failure (HF) poses a significant health burden in high-income countries, affecting over 10% of individuals aged 70 and older. The progressive nature of HF leads to frequent hospital admissions and elevated healthcare costs. There is no easily-acquired biomarker available which allows to gain insight and derive predictions into HF dynamics with the current diagnosis being based on blood tests, X-ray or cardiac ultrasound. Wearable devices like the Apple Watch (Apple Inc., Cupertino, California, U.S.) offer a more unobtrusive monitoring of health parameters compared to these measures. In this feasibility study, we aim to determine in how far Apple Watch data combined with machine learning can be used to predict the course of HF in a well-defined cohort.

        A observational clinical study is conducted at the University Medical Center Göttingen’s Department of Cardiology and Pneumology (IRB Ethics Approval No. 23/2/24; funding via German Centre for Cardiovascular Research, DZHK). N=32 HF patients with reduced ejection fraction hospitalized for acute decompensation (ejection fraction ≤40%, NTproBNP >1000 pg/ml, and at least one symptom such as edema, pleural effusion, or ascites) will be included with a 90 day follow-up period. Patients will wear an Apple Watch during their hospitalization and sensor data (e.g. single-lead electrocardiography, SpO2, Respiration, Step counter) will be acquired. This data is used to predict clinical parameters and risk assessment, e.g. prediction of HF course.

        In this talk we present initial insights of our study with the data that is available at this point. Next to the main research question, we will present results regarding the patients’ opinion on using the Apple Watch and their adherence.

        Speaker: Philip Hempel (Department of Medical Informatics, University Medical Center Göttingen, Göttingen, Germany)
      • 58
        Backpropagation vs. Forward-Forward: unveiling different insights of Duchenne Muscular Dystrophy from multiphoton microscopy imaging data

        In addition to genetic testing, current diagnostic practices for rare neuromuscular diseases involve muscle biopsy analysis, often hindered by subjective interpretation and variability. Advances in artificial intelligence, particularly deep learning, offer promising solutions to automate and enhance diagnostic accuracy by identifying new quantitative, standardized features and phenotypic expressions within biopsy images. These features can be extracted from neural networks using explainable AI tools.

        Backpropagation is the gold standard for training Convolutional Neural Networks (CNNs). However, its forward and backward passes are not biologically plausible given the data flow in the human brain. In December 2022, Prof. Geoffrey Hinton, one of its inventors, proposed an alternative: the Forward-Forward (FF) algorithm, which avoids backward passes by using two forward passes, preventing previous hidden layers from receiving information from subsequent ones. In 2023, we were the first to implement this algorithm on CNNs.

        In this study, we applied FF-trained CNNs to analyze multiphoton microscopy images of muscle biopsies from patients with rare neuromuscular diseases, specifically Duchenne Muscular Dystrophy (DMD). The images included autofluorescence, second harmonic generation, and third harmonic generation signals (SLAM approach).

        We compared the performance of the FF algorithm with backpropagation on the same CNN, reaching 91% and 98% of accuracy on the test dataset, respectively. Using explainable AI tools, specifically class activation maps, we revealed the decision-making processes of the CNNs, showing that FF and backpropagation use different image properties for DMD diagnosis. We validated and then integrated these features by introducing an attention metric, providing a unique and standardized value to quantify pathological characteristics for immediate clinical use.

        Our findings demonstrate the potential of the FF algorithm in delivering new, reliable and interpretable diagnostic insights from biomedical images. This paves the way for its parallel integration with backpropagation-based networks into clinical workflows, enhancing state-of-the-art analysis pipelines by more efficiently exploiting the complex information content available from biological images.

        Speaker: Riccardo Scodellaro (MPINAT)
      • 59
        AI-Driven Adaptive Systems for Knee Rehabilitation: Leveraging Artificial Mental Models for Personalized Patient Support

        In the rapidly changing field of prevention and rehabilitation, adaptive and personalized Artificial Intelligence (AI) systems play a crucial role as facilitators of patient-centric care. These systems, by analyzing comprehensive data on user behavior and situational context, offer support customized to the specific needs of individual patients [1, 2, 3, 4]. However, patients dealing with ailments or recovery from incidents like knee surgeries may face cognitive challenges. These challenges can result in the creation of data that is incomplete, inaccurate, or biased, significantly hampering the patient's ability to make informed decisions, understand complex medical information, or effectively communicate their symptoms and concerns [5, 6, 7, 8]. Recognizing the significant influence of these cognitive limitations, it is essential to leverage AI healthcare systems designed to mitigate these gaps, ensuring care that aligns perfectly with each patient’s distinct needs and conditions [1, 2].
        The exploration of mental models [9] in therapeutic settings and their integration into AI systems introduces a promising path to improve patient care and treatment outcomes. Mental models are cognitive structures that encompass a patient's perceptions and presumptions about their therapy and rehabilitation journey [10, 11, 12, 13, 14]. Existing research highlights the need for accurately capturing and understanding these models, especially in therapeutic and recovery scenarios [15, 16, 17, 18].
        Our research investigates the potential of using meta-representations of patients’ mental models called artificial mental models (AMM) within healthcare AI systems to enhance patient support, especially in making informed choices amidst the uncertainty and discomfort typical in rehabilitation phases. For instance, Carlos is an amateur soccer player recovering from knee surgery and eager to return to the field. His AMM acts as his representative by communicating directly with his physiotherapist to develop a personalized exercise plan that aligns with his goals and recovery status. The AMM schedules these exercises throughout the week, adjusting the intensity based on feedback from Carlos’s wearable devices that track his pain and performance. It also arranges video consultations with his therapist if it detects issues or deviations in his recovery progress, ensuring Carlos remains on the safest and quickest path to full activity.
        In this work, we investigate the elicitation and individualization of AMMs for patients in rehabilitation situations. Large Language Models (LLMs) are used for the generation of AMMs. Training data are acquired by indirect observations, a quantitative study and direct observation. Our research is driven by the following research questions (RQ):
        • RQ1: Can we elicit an individual AMM for a specific patient by using LLMs?
        • RQ2: Are predictions of the AMM regarding expected pain in exercises comparable with expectations of the patient?
        • RQ3: Are predictions of the AMM regarding expected pain in exercises accurate compared with assessment by the therapist?

        For tackling the research questions, we specified a research design according to Design Science Research (DSR) [19, 20] covering a prospective study with two phases: elicitation and individualization. Within each phase, an artifact is created and evaluated [21]. Objective of the elicitation phase is the generation of a discrimination- and bias-free domain-specific basis AMM in the domain of knee rehabilitation. Here, large scale data are acquired on the one hand by a quantitative study (n=150) on personality traits (Big Five [22]) and expected pain in exercises (e.g., squats). On the other hand, indirect observation of patients with knee issues is applied in terms of a twofold scraping approach. Additionally, large corpora of healthcare dialogues are filtered, analyzed, and prepared for model training, e.g., HealthCareMagic [23], ChatDoctor [24], medAlpaca [25]. The resulting artifact – a discrimination- and bias-free domain-specific basis AMM is evaluated in a technical experiment (ablation study). The individualization phase uses this AMM basis model for fine-tuning an AMM for an individual patient. Here, direct observation of a specific patient is used for training the model. Curated data like medication, rehabilitation therapy plans, data on injury, surgical procedure and duration, complications, self-reported pain scales (e.g., Visual Analog Scale (VAS) [26], Wong-Baker FACES Pain Rating Scale [27]) are integrated as well as non-curated data like movement data, sleep, fitness status etc. of the patient. The resulting artifact – an AMM for a specific patient - is evaluated by an action research approach with the patient.

        As result of the elicitation phase, a technically evaluated domain-specific basis AMM in the domain of knee rehabilitation is generated. For building this model, open source LLMs like BLOOM, Mistral7B or Falcon are applied. For ensuring fairness and mitigating bias in the basis AMM, discrimination and bias checks are used. Here a combination of methods is designed, i.e., counterfactual fairness testing [28], adversarial debiasing [29], fairness metrics such as equality of opportunity [30] or demographic parity [31], and interpretability. The basis AMM is used for fine-tuning the AMM for a specific patient in the individualization phase (RQ1). This patient AMM is evaluated in a real-world rehabilitation situation with the patient as part of a research intervention, evaluating its effect on the real-world situation. For the experimental setting an A/B test is planned with focus on anticipation respectively prediction of pain in a therapy unit. Predictions are generated in two variants: (A) by the patient, and (B) by the AMM. For ground truthing, predictions are mirrored with a subsequent actual pain assessment by the patient in a therapy unit. Thus, overlapping of predictions of patient and AMM in concrete cases as well as their accuracy with respect to ground truth are measured (RQ 2 & 3). Contribution of this research is the support of personalized patient care in knee rehabilitation by leveraging LLMs to create and individualize patient-specific AMM for AI systems in healthcare.

        This research emphasizes the role of Artificial Intelligence (AI) systems in enhancing patient-centric care in the field of prevention and rehabilitation. It particularly focuses on the development and implementation of Artificial Mental Models (AMM) within healthcare AI systems in knee rehabilitation. These models are personalized to account for individual patients' mental models regarding their therapy and rehabilitation journeys. In this work, the potential of using Large Language Models (LLMs) to generate and fine-tune these AMMs, aiming for more accurate, bias-free, and personalized patient care, are highlighted.
        By creating systems that better understand and predict patient needs and responses, healthcare providers can offer more targeted and effective interventions in knee rehabilitation. This could lead to quicker recoveries and improved overall patient outcomes. Additionally, the methodological approach of using AMMs could be expanded to other areas of healthcare, suggesting a broad potential impact on how AI is integrated into patient care systems and routine care.

        References:
        1. Khera, R., Butte, A.J., Berkwits, M., Hswen, Y., Flanagin, A., Park, H., Curfman, G., Bibbins-Domingo, K.: Ai in medicine—jama’s focus on clinical outcomes, patient-centered care, quality, and equity. Jama (2023)
        2. Wan, T.T., Wan, H.S.: Predictive analytics with a transdisciplinary framework in promoting patient-centric care of polychronic conditions: Trends, challenges, and solutions. AI 4(3), 482–490 (2023)
        3. Bollos, L.A.C.L., Zhao, Y., Soriano, G.P., Tanioka, T., Otsuka, H., Locsin, R.: Technologies, physician’s caring competency, and patient centered care: A systematic review. The Journal of Medical Investigation 70(3.4), 307–316 (2023)
        4. Li, W., Ge, X., Liu, S., Xu, L., Zhai, X., Yu, L.: Opportunities and challenges of traditional chinese medicine doctors in the era of artificial intelligence. Frontiers in Medicine 10 (2023)
        5. Timm, A., Schmidt-Wilcke, T., Blenk, S., Studer, B.: Altered social decision making in patients with chronic pain. Psychological Medicine 53(6), 2466–2475 (2023)
        6. Meyer, A.N., Giardina, T.D., Khawaja, L., Singh, H.: Patient and clinician experiences of uncertainty in the diagnostic process: current understanding and future directions. Patient Education and Counseling 104(11), 2606–2615 (2021)
        7. Dildine, T.C., Amir, C.M., Parsons, J., Atlas, L.Y.: How pain-related facial expressions are evaluated in relation to gender, race, and emotion. Affective Science pp. 1–20 (2023)
        8. Dildine, T.C., Necka, E.A., Atlas, L.Y.: Confidence in subjective pain is predicted by reaction time during decision making. Scientific reports 10(1), 21373 (2020)
        9. Norman, D.A.: Some observations on mental models, pp. 241–244. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA (1987)
        10. Jones, N.A., Ross, H., Lynam, T., Perez, P., Leitch, A.M.: Mental models: an interdisciplinary synthesis of theory and methods. Ecology and Society 16(1), 46 (2011), https://api.semanticscholar.org/CorpusID:38976887
        11. LaMere, K., M¨antyniemi, S., Vanhatalo, J., Haapasaari, P.: Making the most of mental models: Advancing the methodology for mental model elicitation and documentation with expert stakeholders. Environmental Modelling & Software 124, 104589 (2020)
        12. Borders, J., Klein, G., Besuijen, R.: Mental model matrix: Implications for system design and training. Journal of Cognitive Engineering and Decision Making p. 15553434231226317 (2024)
        13. Johnson-Laird, P.N.: Mental models. The MIT Press (1989)
        14. Hermans, A., Muhammad, S., Treur, J.: You feel so familiar, you feel so different: A controlled adaptive network model for attachment patterns as adaptive mental models. In: Mental Models and Their Dynamics, Adaptation, and Control: A SelfModeling Network Modeling Approach, pp. 321–346. Springer (2022)
        15. Hermans, A., Muhammad, S., Treur, J.: You feel so familiar, you feel so different: A controlled adaptive network model for attachment patterns as adaptive mental models. In: Mental Models and Their Dynamics, Adaptation, and Control: A SelfModeling Network Modeling Approach, pp. 321–346. Springer (2022)
        16. Gabbas, M., Kim, K.: Gamified user interface design for dysphagia rehabilitation based on common mental models. In: DRS2022, Bilbao, Spain (2022)
        17. Barber, T., Crick, K., Toon, L., Tate, J., Kelm, K., Novak, K., Yeung, R.O., Tandon, P., Sadowski, D.C., Veldhuyzen van Zanten, S., et al.: Gastroscopy for dyspepsia: Understanding primary care and gastroenterologist mental models of practice: A cognitive task analysis approach. Journal of the Canadian Association of Gastroenterology 6(6), 234–243 (2023)
        18. Naik, A.D.: Collaborative decision-making: Identifying and aligning care with the health priorities of older adults. In: Geriatric Medicine: A Person Centered Evidence Based Approach, pp. 1–21. Springer (2023)
        19. Hevner, A.R., March, S.T., Park, J., Ram, S.: Design science in information systems research. MIS Quaterly 28, 75–105 (2004), https://api.semanticscholar.org/CorpusID:13553735
        20. Peffers, K., Tuunanen, T., Rothenberger, M.A., Chatterjee, S.: A design science research methodology for information systems research. Journal of Management Information Systems 24, 45 – 77 (2007), https://api.semanticscholar.org/CorpusID:17511997
        21. Peffers, K., Rothenberger, M., Tuunanen, T., Vaezi, R.: Design science research evaluation. In: Design Science Research in Information Systems. Advances in Theory and Practice: 7th International Conference, DESRIST 2012, Las Vegas, NV, USA, May 14-15, 2012. Proceedings 7. pp. 398–410. Springer (2012)
        22. Oliver P. John, Laura P. Naumann, Christopher J. Soto: (2008) Paradigm Shift to the Integrative Big Five Trait Taxonomy. Handbook of Personality Theory and Research. 3. Auflage. S. 114–117
        23. https://huggingface.co/datasets/RafaelMPereira/HealthCareMagic-100k-Chat-Format-en
        24. Li Y, Li Z, Zhang K, et al. (June 24, 2023) ChatDoctor: A Medical Chat Model Fine-Tuned on a Large Language Model Meta-AI (LLaMA) Using Medical Domain Knowledge. Cureus 15(6): e40895. doi:10.7759/cureus.40895
        25. https://github.com/kbressem/medAlpaca
        26. Crichton, N.: Visual analogue scale (vas). J Clin Nurs 10(5), 706–6 (2001)
        27. Wong, D.L., Baker, C.M.: Wong-baker faces pain rating scale. Pain Management Nursing (2012)
        28. Kusner, M.J., Loftus, J., Russell, C., Silva, R.: Counterfactual fairness. In: Advances in Neural Information Processing Systems. vol. 30 (2017).
        29. Lemoine, B., Zhang, B., Mitchell, M.: Mitigating unwanted biases with adversarial learning. In: Proceedings of the ACM Conference (2018)
        30. Hardt, M., Price, E., Srebro, N.: Equality of opportunity in supervised learning. In: Advances in Neural Information Processing Systems. vol. 29 (2016).
        31. Zafar, M.B., Valera, I., Rogriguez, M.G., Gummadi, K.P.: Fairness constraints: Mechanisms for fair classification. In: Proceedings of the 20th International Conference on Artificial Intelligence and Statistics. Proceedings of Machine Learning Research, vol. 54, pp. 962–970. PMLR (2017).

        Speaker: Wolfgang Maaß
      • 60
        Mitigation of privacy issues in the development of clinical AI models

        The development of highly accurate AI models for clinical applications, such as disease classification or personalized treatment recommendation, is heavily constrained by small amount of available training data. The data is commonly generated at different institutions, subject to different data generating processes and privacy regulations. In order to maximize the predictive performance, all of these challenges have to be addressed simultaneously.

        Commonly used Federated Learning (FL) approaches solely aggregate the knowledge of multiple confidential data source by distributing the model development. However, they do not address the high heterogeneity of the different data sources, i.e., due to different experimental equipment, which can introduce significant bias for unseen data.

        To address this issue we introduce Federated Adversarial Cross Training (FACT). FACT uses the implicit domain differences between the clients to identify domain shifts. In each round of FL, FACT cross initializes a pair of clients to generate domain specialized representations which are subsequently used as a direct adversary to learn a domain invariant data representation. We show that FACT not only outperforms state-of-the-art federated, non-federated and source-free domain adaptation models on popular FL benchmarks, but also improves AI model development in clinical applications.

        Speaker: Stefan Schrod (University Medial Center Göttingen)
    • 🇬🇧 Session 6. Multi-Modal Foundation Models Hannah-Vogt-Saal

      Hannah-Vogt-Saal


      Chair: Dr. Joachim Köhler (Fraunhofer IAIS, WestAI)
      Moderation: Laszlo Friedmann (Fraunhofer IAIS, WestAI)

      Content / Abstract:
      With the rapid progress in the field of large language models in recent years, the transfer of the un-derlying technology, i.e. foundation models, to new modalities has become one of the most im-portant research topics in artificial intelligence. With the introduction of CLIP at latest, this develop-ment has been extended to multi-modal foundation models which are able to process different types of modalities, such as images or text. Due to their outstanding properties, such as their excellent ze-ro- or few-shot capability, and their ability to process different modalities, multi-modal foundation models offer huge potential across domains and applications. The overall scope of this session is therefore intended to be broad and to cover all topics related to multi-modal foundation models.

      Topics of interest include:
      - Vision / sound / language / ... models in any possible combination
      - Data- and energy-efficient pre-training
      - Methodologies for efficient transfer and model compression
      - Application to specific domains
      - Ethics, risks, and fairness
      - Securing private data

      • 61
        Open foundation models: reproducible science of transferable learning

        Breakthroughs in strongly transferable learning were achieved by training models that use simple, generic losses and large amounts of generic, diverse web-scale data. Crucial for the progress was increasing pre-training scales, that is model, compute and dataset scales employed in the training. Derived scaling laws suggest that generalization and transferability improve when increasing scales hand in hand. Studying learning at such large scales is challenging, as it requires corresponding datasets at sufficiently large scales to be available, sufficient compute resources to execute the training, while handling properly distributed training across thousands of compute nodes without suffering instabilities. We show how work done by LAION community made the whole pipeline for training strongly transferable multi-modal models of various kind (openCLIP, openFlamingo) - termed foundation models - fully open and reproducible. We show how important experiments necessary for studying such models, for instance those leading to scaling laws derivation, critically depend on the open and reproducible nature of such pipelines - requiring also open-sourcing dataset composition and model benchmark evaluation procedures. We conclude with an outlook on studying next generation of open multi-modal foundation models that possess stronger and more robust generalization, and datasets necessary for their creation.

        Speaker: Dr Jenia Jitsev (Forschungszentrum Jülich)
      • 62
        Importance of data in multi-modal modelling
        Speaker: Marianna Nezhurina (Forschungszentrum Jülich)
      • 63
        High Fidelity Video Background Music Generation with Transformers

        The recent success of large language models (LLMs) like GPT and BERT has demonstrated the immense capabilities of transformer-based architectures on natural language processing (NLP) tasks, such as text generation, translation, and summarization, setting new benchmarks in Artificial Intelligence (AI) performance. Building on this momentum, the AI research community is increasingly focusing on extending the capabilities of LLMs to multi-modal data, giving rise to multimodal foundation models. The use of generative models for music generation has also been gaining popularity. In this study, we present the application of multi-modal foundation models in the domain of video background music generation. Current music generation models are predominantly controlled by a single input modality: text. Video input is one such modality, with remarkably different requirements for the generation of background music accompanying it. Even though alternative methods for generating video background music exist, none achieve a music quality and diversity that is comparable to the text-based models. We adapt the text-based models to accept video as an alternative input modality for the control of the audio generation process and evaluate our approach quantitatively and qualitatively through the analysis of exemplary results in terms of audio quality and through a case study to determine the users’ perspective on the video-audio correspondence of our results.

        Speaker: Yongli Mou (RWTH Aachen University)
      • 64
        Multi-Modal Foundation Models in Production – Potentials and Challenges

        The emergence of multi-modal foundation models, i.e., large AI models pre-trained on vast amounts of data of different modalities, which show emergent behavior and generalization ability over a set of different tasks, brings enormous possibilities across industries. Specific use cases in the production sector around this technology are however currently scarcely available. Therefore, this presentation gives an overview about the possibilities and challenges of such models for productional use cases.

        Speaker: Hannes Behnen
      • 65
        Enhancing Transparency and Trustworthiness in Neuro-Symbolic Scholarly Knowledge Curation Using Language and Visual Models

        Scholarly knowledge curation encounters challenges due to the diverse methodologies across scientific fields. Tailored approaches and protocols are essential to address these challenges, considering each domain's unique characteristics. Machine assistance, particularly through Large Language Models (LLMs) such as GPT-3.5, offers significant potential to navigate these complexities and enhance knowledge curation processes. The Open Research Knowledge Graph (ORKG) is a structured semantic platform to mitigate issues associated with traditional document-centric scholarly communication, including publication overload and peer review deficiencies.
        The rapid evolution of multi-modal foundation models in artificial intelligence has transformed the integration of diverse modalities, including language and visual data. In this session, we aim to explore innovative approaches that leverage these multi-modal models alongside prompt engineering techniques. Emphasizing explainability, transparency, and trustworthiness, our research focuses on the curation of scholarly knowledge. By integrating LLMs (e.g., GPT-3.5) and visual models, we aim to deepen insights and enhance the accessibility of knowledge extracted from academic literature.
        In this session, we demonstrate our approach using LLMs (e.g., GPT-3.5) to extract data from research articles, facilitating innovative insights and accelerating research outcomes. Our method integrates LLM capabilities with ORKG, validated by domain experts to ensure data accuracy and relevance. This collaborative framework merges neural LLM capabilities with symbolic knowledge graphs and human expertise, addressing practical challenges in applying LLMs to scientific research. We leverage both language and visual models to enrich our understanding and interpretation of scholarly content, thereby broadening the scope of insights generated.
        We employ prompt engineering with predefined templates to effectively interact with LLMs, ensuring consistent and precise data extraction. Our discussion will cover strategies for leveraging LLMs to generate novel insights, challenges encountered in their application, and solutions developed to overcome these obstacles. Emphasis will be placed on tools for enhancing AI explainability and trustworthiness, ensuring transparent and reliable use of LLMs in scientific research.

        Speaker: Hassan Hussein (LUH)
    • 🇬🇧 Session 7. Applications of artificial intelligence in the energy sector Emmy-Noether-Saal

      Emmy-Noether-Saal


      Chair: André Baier, Dominik Beinert (Fraunhofer IEE)

      Content / Abstract:
      This session focuses on the application of artefitial intelligence in the energy sector. We are looking for talks from various related topics which explore the challenges and some of the employed solutions for the critcal energy infrastructure.

      Topics of interest include:
      - Forecasting generation and consumption
      - State estimation for electrical grids
      - Vegetation detection for overhead condutors
      - Bird detection on wind sites
      - Autonomous agents for grid operation und energy markets

      Agenda
      - Welcome: 5 min
      - 4 contributed talks (each 15 min + 5 min discusison)
      - Conclusion: 5 min

      • 66
        Remote sensing driven machine learning methods in the field of energy system analysis

        The expansion of renewable energies is of crucial importance for the future of energy supply. There are numerous reasons for this change and highlight the need for a transition from fossil fuels to sustainable energy sources. Access to clean energy plays a central role in improving the quality of life and protecting our environment. A diverse data basis is essential for efficient implementation at the various levels.

        In recent years, the number of satellites orbiting the earth has increased rapidly. They are used for a variety of applications. For example, satellite based services include global communication via. Mobile telephony and internet connections, navigation using GPS, weather forecasts based on weather data and the provision of up-to-date maps of the earth's surface for various mapping services are part of the daily habits of billions of people. Favorable drivers of advancing technologization and decreasing costs for satellite launches have led to an increasing number of different players such as countries and companies being able to operate satellites in Earth orbit. This results in large quantities of data collected by satellites, which is often made available at low cost or even free of charge. However, this is also accompanied by the great challenge of handling such enormous amounts of data. Methods from the field of machine learning can provide a remedy for handling the enormous amounts of data that arise in the field of remote sensing.

        Machine learning methods offer a wide range of possible applications in data driven domains such as the financial industry, the automotive industry and also in the various areas of image-based processes. For example, recurring and conspicuous patterns can be identified in large data sets. Large amounts of data collected by satellites on land use, soil conditions, cover and other geographical features can be processed automatically using machine learning methods to obtain information that would otherwise be difficult or impossible to obtain.

        When the two areas of remote sensing and machine learning are combined, a wide range of possibilities arise for various aspects of the energy system. For example, machine learning methods can be combined with aerial and satellite images to map the stock of renewable energy plants. Various studies have already dealt with the automated detection of PV, wind power, biogas and hydropower plants. The methods and data used vary greatly. In the planned presentation, 3 different studies on remote sensing-based machine learning for energy system analysis will be presented. The resulting findings will then be used to identify paths for the energy system landscape.

        • The first article shows the extent to which semi-automated approaches can be used to create training patterns for object recognition in photovoltaic systems on roofs. It is shown that the time and effort
        required for manual identification can be greatly reduced if existing spatial information is combined automatically. The article has already been published (Kleebauer et al. 2021). The methodology can also be used in a similar way to improve the location accuracy of wind turbines (Kleebauer et al. 2024).

        • In a further contribution, varying training data will be examined with regard to their suitability for the detection of renewable energy plants. Various modern deep learning-based approaches for segmentation
        will be combined and compared with a wide variety of aerial images and satellite images. (Kleebauer et a. 2023).

        • In a third contribution, methods for image sharpening in satellite and aerial images will be investigated using various deep learning methods. Various popular architectures such as Generative Adversarial Networks and various Transformer Networks can be used here (Horst & Kleebauer 2022).

        The results of the various contributions are then summarized and examined with regard to their future suitability, associated challenges and further potential in the area of the energy system landscape.

        Speaker: Maximilian Kleebauer (Department of Energy Management and Power System Operation, University of Kassel; Energy Meteorology and Geo Information System, Fraunhofer IEE, Kassel)
      • 67
        Optimizing Renewable Energy Forecasts: An MLOps Conceptual Approach for Scalability

        As the expansion of renewable energies grows, the need for accurate energy forecasts becomes crucial due to the dependency on volatile energy sources. Traditional forecasting systems, which utilize weather data and historical generation data, are challenged by the unique behaviors of individual power plants, lack of data and changing conditions (Yan et al. 2022). To address these challenges, we propose a highly scalable energy prediction system based on MLOps principles to ensure efficient model updates while adhering to appropriate quality criteria. This aims to improve the accuracy and efficiency of renewable energy forecasts, supporting a reliable energy supply in the dynamic energy sector.
        The primary purpose of MLOps is to efficiently facilitate the deployment of ML models into production by eliminating bottlenecks in Development and Operations and automating the workflows (Subramanya et al. 2022). Building on these principles, we aim to develop a service mesh comprising multiple interacting microservices, which include a training service for both experimental and production environments, a forecasting service for operational forecasts, a centralized feature store that houses all necessary data including training, test, operational data, and master data, a model store for archiving models with their parameters and metrics, and a monitoring service to ensure prediction quality and service reliability.
        The main challenge we address is the large number of assets, each requiring individual model predictions, necessitating a highly flexible and scalable ML pipeline. We employ dual training modes for initial and continuous model retraining within the same productive Kubernetes cluster used for inference. Furthermore, a separate model store for logging and tracking supports access at any stage of the ML pipeline (Alla and Adari 2021), complemented by a central feature store that enhances flexibility and adaptability by utilizing consistent data interfaces (Dowling 2023). For a high scalability of our inference service, we use Horizontal Pod Autoscaling provided by Kubernetes that automatically updates a workload resource to match request demand (The Kubernetes Authors 2024).
        In summary, conventional static methods do not fulfill the growing requirements of increasingly frequent dynamic changes in modern energy systems. We therefore aim to use MLOps concepts to provide scalable services for the continuous improvement of forecasts under changing conditions.

        Literature
        Alla, Sridhar; Adari, Suman Kalyan (2021): Introduction to MLFlow. In Sridhar Alla, Suman Kalyan Adari (Eds.): Beginning MLOps with MLFlow. Berkeley, CA: Apress, pp. 125–227.
        Dowling, John (2023): What is a Feature Store for Machine Learning? Available online at https://www.featurestore.org/what-is-a-feature-store, updated on 8/11/2023, checked on 5/17/2024.
        Subramanya, Rakshith; Sierla, Seppo; Vyatkin, Valeriy (2022): From DevOps to MLOps: Overview and Application to Electricity Market Forecasting. In Applied Sciences 12 (19), p. 9851. DOI: 10.3390/app12199851.
        The Kubernetes Authors (2024): Horizontal Pod Autoscaling. Available online at https://kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/, updated on 2/18/2024, checked on 5/17/2024.
        Yan, Jie; Möhrlen, Corinna; Göçmen, Tuhfe; Kelly, Mark; Wessel, Arne; Giebel, Gregor (2022): Uncovering wind power forecasting uncertainty sources and their propagation through the whole modelling chain. In Renewable and Sustainable Energy Reviews 165, p. 112519. DOI: 10.1016/j.rser.2022.112519.

        Speakers: Raphael Riege (Fraunhofer IEE), Johannes Schütz (Fraunhofer IEE)
      • 68
        CARE to Compare: A real-world benchmark dataset for early fault detection in wind turbine data

        Early fault detection is crucial in predictive maintenance for wind turbines, yet comparing different algorithms remains challenging due to the scarcity of domain-specific public datasets. Many papers introduce sophisticated algorithms based on inaccessible data, making their results difficult to verify and hard to compare with results of similar algorithms. This presentation addresses these issues by introducing a newly published, high-quality dataset that includes data from 36 wind turbines across three wind farms, offering the most detailed fault information available for public wind turbine datasets.
        Additionally, we propose a novel scoring method, CARE (Coverage, Accuracy, Reliability, and Earliness), designed to leverage the dataset's depth to evaluate early fault detection models effectively. This method assesses anomaly detection performance, the ability to recognize normal behavior, and the capability to minimize false alarms while detecting anomalies early.

        Speaker: Cyriana Roelofs (Fraunhofer IEE, Kassel)
      • 69
        KPI Extraction from Maintenance Work Orders—A Comparison of Expert Labeling, Text Classification and AI-Assisted Tagging for Computing Failure Rates of Wind Turbines

        Maintenance work orders are commonly used to document information about wind turbine operation and maintenance. This includes details about proactive and reactive wind turbine downtimes, such as preventative and corrective maintenance. However, the information contained in
        maintenance work orders is often unstructured and difficult to analyze, presenting challenges for decision-makers wishing to use it for optimizing operation and maintenance. To address this issue,
        this work compares three different approaches to calculating reliability key performance indicators from maintenance work orders. The first approach involves manual labeling of the maintenance work
        orders by domain experts, using the schema defined in an industrial guideline to assign the label accordingly. The second approach involves the development of a model that automatically labels
        the maintenance work orders using text classification methods. Through this method, we are able to achieve macro average and weighted average F1-scores of 0.75 and 0.85 respectively. The third technique uses an AI-assisted tagging tool to tag and structure the raw maintenance information, together with a novel rule-based approach for extracting relevant maintenance work orders for failure rate calculation. In our experiments, the AI-assisted tool leads to an 88% drop in tagging
        time in comparison to the other two approaches, while expert labeling and text classification are more accurate in KPI extraction. Overall, our findings make extracting maintenance information from maintenance work orders more efficient, enable the assessment of reliability key performance indicators, and therefore support the optimization of wind turbine operation and maintenance.

        Speaker: Bastian Schäfermeier (GNOI)
    • 🇬🇧 Workshop 3. Data Science & Research Data Taberna

      Taberna



      Instructurs: Prof. Sören Auer, Dr. Markus Stocker, Dr. Oliver Karras (all University of Hannover LUH)

      Agenda:

      • Presentation of the Open Research Knowledge Graph for the organisation of research information
      • Organisation of research data with Jupyter Notebooks
      • Organisation of Energy Scenario Factsheets in the ORKG
      • Ontology-based research data management

      Target Group: Researchers in the field of AI medicine and energy, users of KISSKI and those interested in research data management

      Prerequisites: None. A laptop ist usful, but not required.

      Max. number of participants: 20

    • 1:15 PM
      Mittagspause / Lunch Break

      Oberes Foyer / Upper Foyer & Adam-von-Trott-Saal

    • 🇩🇪 Podiumsdiskussion "AI made in Germany: Wie beschleunigen KI-Servicezentren den Transfer von KI-Anwendungen in Unternehmen? " Adam-von-Trott-Saal

      Adam-von-Trott-Saal


      Zusammensetzung des Podiums

      • Moderation: Holger Karl (Hasso-Plattner-Institut für Digital Engineering)
      • Florian Kieser (HessianAI)
      • Marie-Jolin Köster (Deutsche Krebsgsellschaft e.V., Berlin)
      • Michael Markeev (Deutsche Bank)
      • Maximilian Carlos Menke (My Taste)
      • Pina Merkert (Technikjournalistin)
      • Cornelius Scheffel (u-form Verlag und u-form Testsysteme)

      Über die Mitglieder des Podiums

      Prof. Dr. Holger Karl (Akademischer Leiter von AI Services, dem KI-Servicezentrum Berlin-Brandenburg) – Moderator des Panels
      Holger Karl ist derzeit Professor für Informatik am Hasso-Plattner-Institut und beschäftigt sich mit Netzwerksoftware, mobilen und drahtlosen Netzwerken sowie Rechenzentren. Er hat mehrere Forschungsprojekte am HPI geleitet, darunter das open6GHub-Projekt und das KI-Servicezentrum Berlin-Brandenburg. Seine akademische Laufbahn begann mit einem Abschluss in Informatik am Karlsruher Institut für Technologie, gefolgt von einer Promotion an der Humboldt-Universität zu Berlin und einem Postdoc-Aufenthalt an der Technischen Universität Berlin. Von 2004 bis 2021 war er Professor an der Universität Paderborn in der Computer Networks Group. Seine Forschung bewegt sich an der Schnittstelle von Netzwerken, verteilten Systemen und maschinellem Lernen.

      Florian Kieser (hessianAI)
      Florian Kieser ist Leiter vom hessischen KI-Servicezentrum und stellvertretender CTO von hessianAI. Mit seiner 13-jährigen Erfahrung als Projektleiter in verschiedenen Wirtschaftsunternehmen und einer eigenen Start-up-Gründung verfolgt er das Ziel, den Zugang und die Anwendung von KI für jeden so einfach wie möglich zu gestalten.

      Marie-Jolin Köster (Deutsche Krebsgsellschaft e.V., Berlin)
      Marie-Jolin Köster ist Anthropologin und leitet seit 2017 die Abteilung Wissensmanagement | INFONETZ KREBS der Deutschen Krebsgesellschaft e.V.. Die Abteilung erstellt evidenzbasierte Informationen zum Thema Krebs sowohl in Fachsprache als auch in allgemeinverständlicher Form. Das erste KI-Projekt startete sie im Jahr 2017 mit dem Ziel, maschinelles Lernen zur Bewältigung großer Datenmengen im Zusammenhang mit systematischen Literaturrecherchen einzusetzen. Derzeit arbeitet die Abteilung an einer KI-Anwendung zur Unterstützung der Textredaktion.
      Schwerpunkte der Arbeit von Frau Köster sind: Organisation von Wissen sowie Innovations- & Change-Management.

      Michael Markeev (Deutsche Bank)
      Michael Markeev arbeitet im Innovations-Team der Deutschen Bank in Berlin und verantwortet dort einige Projekte zur Erprobung generativer KI-Technologien im Bereich „Conversational AI“. Als Gründungsmitglied der bankinternen AI/ML Guild Berlin engagiert er sich in der Förderung und Vernetzung der internen KI-Community. Zuvor war er seit 2021 als AI Product Lead tätig und leitete mehrere Innovationsprojekte in unterschiedlichen Bereichen der Bank. Im Rahmen seiner Tätigkeit für das Innovationsteam arbeitete er intensiv mit Start-ups zusammen, um neueste Technologien aus den Bereichen KI, VR/AR und Data Privacy Engineering einzuführen.

      Maximilian Carlos Menke (MyTaste)
      Maximilian Carlos Menke studiert Physik an der RWTH Aachen. Seine erste Informatikerfahrung sammelte er bereits während der Schulzeit durch die Entwicklung einer Online-Plattform für einen Sportwissenschaftler, die zur individuellen Analyse von Leistungsdaten genutzt wurde. Anschließend arbeitete er für das Start-up Graswald aus Hannover, wo er an der Entwicklung des fortschrittlichsten Scanners für die Erstellung von 3D-Assets von Pflanzen mitwirkte. Seit 15 Monaten arbeitet er an dem Start-up MyTaste, das mit Unterstützung der Rechenressourcen des KI-Servicezentrums an der Weiterentwicklung von Bildgenerierungsmodellen forscht.

      Pina Merkert (Technikjournalistin)
      Pina Merkert hat nach 4 Semestern Informatik in Kaiserslautern in Mainz Mediendramaturgie studiert. Nach Berufserfahrung als Webentwicklerin in Kaiserslautern und Tokyo arbeitete sie von 2017 bis 2024 als Technikjournalistin beim c't-Magazin in Hannover. Neben Softwareentwicklung und 3D-Druck beschäftigte sie sich in ihren Artikeln vor allem mit neuronalen Netzen. Dabei begnügte sie sich nicht damit, die Entwicklung zu begleiten, sondern veröffentlichte zahlreiche praktische Anleitungen, um KI-Technik selbst nachzubauen und zu trainieren.

      Cornelius Scheffel (u-form Verlag und u-form Testsysteme)
      Cornelius Scheffel blickt auf eine fast 40-jährige IT-Karriere zurück. Angefangen an Großrechnern über Apple II, PC und alles, was danach kam, war die darauf laufende Software und -architektur sein Steckenpferd. Nach anfänglichen vielen Jahren im Bereich der ERP-Softwarelösungen, hat er sich seit 2008 den Softwaresystemen rund um die duale Ausbildung verschrieben. Zu seinen Aufgabengebieten gehören neben dem aktuellen KI-Projekt mit WestAI die Softwarearchitektur und -entwicklungsprozess sowie der Betrieb der Lösungen in den Rechenzentren der u-form Gruppe.

    • 🇬🇧 Workshop 3. Data Science & Research Data: Continued Taberna

      Taberna



      Instructurs: Prof. Sören Auer, Dr. Markus Stocker, Dr. Oliver Karras (all University of Hannover LUH)

      Agenda:

      • Presentation of the Open Research Knowledge Graph for the organisation of research information
      • Organisation of research data with Jupyter Notebooks
      • Organisation of Energy Scenario Factsheets in the ORKG
      • Ontology-based research data management

      Target Group: Researchers in the field of AI medicine and energy, users of KISSKI and those interested in research data management

      Prerequisites: None. A laptop ist usful, but not required.

      Max. number of participants: 20

    • 🇩🇪 🇬🇧 Industrieausstellung & Networking / Industrial Exhibition & Networking Emmy-Noether-Saal

      Emmy-Noether-Saal

      • 70
        AI Grid Netzwerk für KI-Nachwuchsforscher:innen: Wissenstransfer, Verbindungen, Zusammenarbeit.

        www.ai-grid.org

      • 71
        AI Village

        Am Stand stellt sich das AI Village als Innovationscampus für KI und Robotik in Hürth vor inklusive der verschiedenen kostenfreien Angebote, die es für KI-Interessierte, öffentliche Einrichtungen und Unternehmen aus dem Rheinischen Revier anbietet.

        https://ai-village.eu/

      • 72
        DDN - The AI Data Company

        DDN is a leading global provider of data storage and data management solutions at scale. We accelerate AI and High Performance Computing workflows and applications in data centers, private and public clouds, and at the edge. Thanks to our technology, over 11,000 customers realize significant efficiencies in their GPU and CPU compute farms, substantially reducing their data center power consumption and footprint. Utilizing highly optimized flash technology and AI-enabled software, our products power some of the largest and most demanding customers in the world in fields such as autonomous driving, AI chatbots, healthcare, financial services, manufacturing, energy, government, public sector, and research institutions, as well as generative AI and data analytics applications. Explore our offerings further at ddn.com.

      • 73
        JELLYSPACE (Supply Chain AI-Platform)

        Demo of our Platform on Laptop and explanation of the entire platform with a Pull-up board explanation.
        https://jellyspace.ai

      • 74
        NFDI - Research Data Management and AI

        Joint display of NFDI4ing, NFDI4DataScience, NFDI4energy and NFDI4health. We will show how AI applications can profit from good research data management and how research data management can in turn profit from AI assistance.

        https://www.nfdi.de/

      • 75
        semares - semantic data integration solutions for AI model development and application

        Semares is a data integration platform solution developed by Genevention GmbH. Semares facilitates integration, analysis and exploration of life science data and provides programmatic as well as push-button access for bioinformatics and AI applications.

        www.genevention.com

      • 76
        The Data Platform for the Cloud & AI Era

        https://www.weka.io/

      • 77
        Vertrauenswürdige KI mit Confidential Computing

        Wir werden unser Produkt Continuum AI ausstellen, das als erstes Nutzerdaten Ende-zu-Ende mit Hilfe der Confidential Computing Features der Nvidia H100 schützt. Continuum AI wird bald als Open-Source veröffentlicht.

        Edgeless Systems GmbH
        www.edgeless.systems

    • 🇩🇪 🇬🇧 Industrieausstellung & Networking / Industrial Exhibition & Networking Hannah-Vogt-Saal

      Hannah-Vogt-Saal

      • 78
        Hessian.AI Servicezentrum für künstliche Intelligenz / Hessian.AI Service Center for Artificial Intelligence

        https://hessian.ai/de/ki-servicezentrum/

      • 79
        KI für eine personalisierte Medizin / AI for personalized medicine

        Wie / wo können KI- und kausale Methoden zu einer verbesserten, personalisierten Gesundheitsversorgung beitragen.

        CAIMed - Lower Saxony Center for AI and Causal Methods in Medicine
        www.caimed.de

      • 80
        KI-Servicezentrum Berlin-Brandenburg / AI Service Center Berlin-Brandenburg

        https://hpi.de/kisz

      • 81
        KISSKI: KI-Servicezentrum für sensible und kritische Infrastrukturen / KISSKI: AI-Servicecenter for Critical and Sensitive Infrastructure

        https://kisski.gwdg.de

      • 82
        mlguide: Methodological Guidance for Applied Machine Learning

        Our software demonstrator mlguide facilitates the selection of suitable methods for each step of the ML workflow. It is based on expert knowledge from the literature and comes with a user-friendly, interactive frontend. We can show how mlguide is successfully integrated into our visual Machine Learning Toolbox. Furthermore, we can give an overview on AI in Digital Medicine at Fraunhofer MEVIS, Bremen.

        Fraunhofer-Institut für Digitale Medizin MEVIS
        https://www.mevis.fraunhofer.de/

      • 83
        ScaDS.AI Dresden/Leipzig

        Current activities of ScaDS.AI as one of the 6 federal AI centers, and its Service Center.

        scads.ai

      • 84
        WestAI-Servicezentrum / KI Service Center WestAI

        https://westai.de/

    • 🇩🇪 Zusammenfassung & Abschluss / Summary & Good-Bye Adam-von-Trott-Saal

      Adam-von-Trott-Saal

      Convener: Prof. Julian Kunkel (GWDG / Georg-August-Universität Göttingen)