scieee AI-readable full text Open interactive document viewer

Data Infrastructures as a Foundation for AI Projects

Schimmler, Sonja

Publisher
Zenodo
DOI
10.5281/zenodo.17878884
Repository
zenodo.org
Language
en

Abstract

This presentation will examine the critical components that drive AI progress, with a focus on the need for large-scale, machine-interpretable data that adapts to the unique demands of different sectors. Key topics will include transparency and reproducibility in AI, as well as the importance of making essential resources—publications, data, models, and code—accessible, interconnected, and scalable. A robust data infrastructure is essential for the AI community, supporting every stage of the data lifecycle—from collection and creation to processing, analysis, publishing, archiving, and reuse of resources. This talk will highlight the foundational role of these infrastructures in AI-driven initiatives, showcasing several ongoing infrastructure initiatives and illustrating how they contribute to a modern data ecosystem.

Full text

Data Infrastructures and Data Competencies as a Foundation for AI Projects Prof. Dr. Sonja Schimmler TU Berlin, Fraunhofer FOKUS and Weizenbaum Institute DiTraRe Symposium 2025 Motivation: AI transforming the scientific process across disciplines 202.12.2025, Sonja Schimmler Fraction of LLM-modified sentences in publications over time [Liang et al., Nature Hum Behav 2025] ?? ? ➔AI in data science & analysis pipelines Data analysis ➔AI adoption in research infrastructures Searching and integrating metadata, papers or research data ➔AI for labeling and generating data Simulating humans and synthetic data ➔AI assisting researchers in the scientific process Writing code or reviewing papers … ➔AI (semi-)autonomously conducting science? Motivation: AI transforming the scientific process across disciplines 3 AI Evolution Deep learning Representation learning Foundation models & LLMs Transfer learning Retrieval-augmented generation Agentic AI … AI Adoption in Science 02.12.2025, Sonja Schimmler ● Robustness & generalisability Shortcut learning & benchmark leakage; e.g. “predicting pneumonia” from x-rays [Geirhos et al., Nature Mach Intell 2020] ● Reproducibility & state-of-the-art crisis E.g. only 38% of top-tier deep learning papers are reproducible and only 5% beat simple baselines [Dacrema et al., RecSys 2019] ● Bias & discrimination E.g. health risk of people of color consistently underestimated by predictive algorithms. [Obermeyer et al., Science 2019] Motivation: Critical challenges for AI adoption in science 402.12.2025, Sonja Schimmler NFDI4DS: NFDI and NFDI4DS 502.12.2025, Sonja Schimmler NFDI for Data Science and Artificial Intelligence ●NFDI to set up a German National Research Data Infrastructure The goal is to build ONE NFDI NFDI4DS: A national research data infrastructure for data science and AI ● NFDI4DS is setting up a … ○ … National Research Data Infrastructure for Data Science and Artificial Intelligence ● Our Goals ○ Develop, establish and sustain a national research data infrastructure and offer innovative tools and services ○ Make all digital artefacts (articles, data, models, workflows, scripts/code) available ○Interlink all digital artefacts 6 DO Metadata Service Interfaces Identifier Knowledge Graphs (KGs) FAIR Digital Objects (FDOs) Neuro-symbolic Methods 02.12.2025, Sonja Schimmler NFDI4DS: Towards adoption of AI in science 7 AI assistance in the whole AI research lifecycle FAIRness of paper, data, models, scripts/code, … AI literacy education & training, guidelines & best practices Reproducibility of science Responsible use of AI across disciplines Transparency of data & models (bias & discrimination) Fair benchmarking of models (robustness & generalisability) 02.12.2025, Sonja Schimmler Infrastructure and services - 4DS meta portal 802.12.2025, Sonja Schimmler ●Utilizing knowledge graphs and AI methods Resources Shared tasks - Readme2KG 902.12.2025, Sonja Schimmler 4DS Ontology GitHub ReadMe Files ●Development of new tools and services ●Community involvement