Welcome

• Claus Trost (Erich Schmid Institute of Materials Science of theAustrian Academy of Sciences)
• Jan Odstrčilík (Institut für Mittelalterforschung, ÖAW)
• Kati Heinrich (IGF | ÖAW)
• Nicki Holighaus (Acoustics Research Institute, Austrian Academy of Sciences)
• Claudius Krause (MBI Vienna (ÖAW))

Room: Seminar room 1 + 2

What is AI? Overview lecture: Part I

• Michal Hradiš (Brno University of Technology)

Room: Seminar room 1 + 2 1) This lecture introduces what “AI” means today and how it relates to machine learning, with an emphasis on large language models (LLMs) and AI agents. We’ll build intuition for core ML ideas—supervised vs. self-supervised learning, loss functions, optimization, and generalization. regression, classification, generation, language models, etc.

AI and ML at the Austrian Academy of Sciences

• Claudius Krause (MBI Vienna (ÖAW))

Room: Seminar room 1 + 2

Overview lecture: Part II

• Michal Hradiš (Brno University of Technology)

Room: Seminar room 1 + 2 2) This lecture explains how today’s language models work at a high level: key building blocks, typical training stages, and the resulting capabilities and limitations. We’ll finish with a brief Python demo using the OpenAI API and a local model via Ollama. LLMs and LLM architecture, transformers, supervised and unsupervised training in more detail

Vibe Coding & „Promptotyping”: Part I

• Christopher Pollin (DH Craft Graz)

Room: Seminar room 1 + 2

Vibe Coding & „Promptotyping”: Part II

• Christopher Pollin (DH Craft Graz)

Room: Seminar room 1 + 2

Vibe Coding & „Promptotyping”: Part II

• Christopher Pollin (DH Craft Graz)

Room: Seminar room 1 + 2

Vibe Coding & „Promptotyping”: Part III

• Christopher Pollin (DH Craft Graz)

Room: Seminar room 1 + 2

Bias in Machine Learning

• Philip Winter (VRVis)

Room: Seminar room 1 + 2

OpenEuroLLM and other developments on AI infrastructures

• Patrick Rarivoson

Room: Seminar room 1 + 2

AI Strategy at the Austrian Academy of Sciences (ÖAW)

• Bernd Wachmann (ÖAW)

Room: Seminar room 1 + 2

Governing Machines: From Moderation to Alignment

• Charlotte Spencer-Smith (CMC)

Room: Seminar room 1 + 2

AI Policy at the Vienna University

• Laura Gandlgruber (Vienna University)

Room: Seminar room 1 + 2

From idea to final project: Part I - Data & Code Management

• Stavrina Dimosthenous (University of Manchester)

Room: Seminar room 1 + 2 Delving into the world of data-driven research presents exciting possibilities. However, there are a few topics that need to be addressed before one can jump right in. Research data management (RDM) planning, software version control, automation, and data cleaning might appear dry at first glance, and they might indeed be to many. Despite these topics and practices seeming like ‘checkbox’ activities, expending some time actively thinking about them, and setting them up can save a lot of time and set you up for success in your research endeavour. This workshop aims to introduce you to these topics. By the end of the workshop you will be prepared to independently self-teach these topics. We will begin by exploring some academic research project examples and how RDM and RDM planning could be approached. Moving on, I will introduce software version control via Git, with remotes on the developer platform, GitHub, and touching on working as a team on GitHub. This exercise will also explore automation via GitHub Actions and Docker. We finish with data cleaning and wrangling concepts, and a practical demonstration in Python. Participation by following along is encouraged but not expected. Github management, workflows, exercises [https://stavrina.github.io/oeaw-winter-school/setup/][1] [1]: https://stavrina.github.io/oeaw-winter-school/setup/

From idea to final project: Part II - Data & Code Management

• Stavrina Dimosthenous (University of Manchester)

Room: Seminar room 1 + 2 Delving into the world of data-driven research presents exciting possibilities. However, there are a few topics that need to be addressed before one can jump right in. Research data management (RDM) planning, software version control, automation, and data cleaning might appear dry at first glance, and they might indeed be to many. Despite these topics and practices seeming like ‘checkbox’ activities, expending some time actively thinking about them, and setting them up can save a lot of time and set you up for success in your research endeavour. This workshop aims to introduce you to these topics. By the end of the workshop you will be prepared to independently self-teach these topics. We will begin by exploring some academic research project examples and how RDM and RDM planning could be approached. Moving on, I will introduce software version control via Git, with remotes on the developer platform, GitHub, and touching on working as a team on GitHub. This exercise will also explore automation via GitHub Actions and Docker. We finish with data cleaning and wrangling concepts, and a practical demonstration in Python. Participation by following along is encouraged but not expected. Github management, workflows, exercises

From idea to final project: Part III - Data cleaning, reproducability

• Stavrina Dimosthenous (University of Manchester)

Room: Seminar room 1 + 2 Delving into the world of data-driven research presents exciting possibilities. However, there are a few topics that need to be addressed before one can jump right in. Research data management (RDM) planning, software version control, automation, and data cleaning might appear dry at first glance, and they might indeed be to many. Despite these topics and practices seeming like ‘checkbox’ activities, expending some time actively thinking about them, and setting them up can save a lot of time and set you up for success in your research endeavour. This workshop aims to introduce you to these topics. By the end of the workshop you will be prepared to independently self-teach these topics. We will begin by exploring some academic research project examples and how RDM and RDM planning could be approached. Moving on, I will introduce software version control via Git, with remotes on the developer platform, GitHub, and touching on working as a team on GitHub. This exercise will also explore automation via GitHub Actions and Docker. We finish with data cleaning and wrangling concepts, and a practical demonstration in Python. Participation by following along is encouraged but not expected.

Introduction to CLIP

• Erich Birngruber (CLIP)

Room: Seminar room 1 + 2

Explainability of Neural Networks in STEM fields

• Gunnar König

Room: Seminar room 1 + 2 Due to their predictive accuracy, machine learning models are increasingly employed in high-stakes decision making and in scientific contexts. These predictive capabilities often come at the price of interpretability: The models are too complex to be inherently intelligible by their human users. In an attempt to restore this interpretability, a broad range of methods have been developed, often subsumed under the umbrella of the term Explainable AI (XAI). In this session, we will give the attendees an introduction to the field. We will start by mapping the different approaches using the standard taxonomies, before we formally introduce some of the most popular explanation methods. Over the course of the session we will focus on the core challenges that practitioners face: (1) Translating the outputs of XAI methods into concise statements about model and data, and (2) choosing the explanation technique that is most suitable for a given task. We will tackle these questions with a particular focus on the task of learning about feature target associations, thereby preparing attendees to employ XAI methods in their own work.

Explainability of Neural Networks in STEM fields

• Moritz Grosse-Wentrup

Room: Seminar room 1 + 2 Machine learning (ML) models are increasingly deployed in high-stakes environments, e.g., in the health domain, where ethical and legal considerations require models to be interpretable. Despite substantial progress in interpretable ML (IML), several key challenges remain. These include distinguishing between interpreting the model and using the model to interpret the data-generating process, dealing with heterogeneous data, and providing meaningful recourse recommendations. I argue that we can address these challenges by adopting a causal perspective on IML. In particular, I demonstrate how the causal structures of the model and the data-generating process jointly affect model interpretation techniques. I then provide an overview of our recent work on leveraging causal techniques to improve model interpretation and render recourse recommendations more meaningful.

Evaluation and Interpretability of NLP systems

• Pia Sommerauer (Vrije Universiteit Amsterdam)

Room: Seminar room 1 + 2