Inhalt

Abstract:

AI systems are increasingly making decisions with real-world consequences – in hiring, credit lending, medical diagnostics, and law enforcement. This course teaches how bias arises in such systems, how it can be measured, and which strategies can reduce it – in both classical machine learning models and large language models (LLMs). After a compact review of the fundamentals of machine learning, the course introduces the concepts of bias and fairness and presents various, at times conflicting, fairness definitions. The causes of bias are then systematically examined along the entire ML lifecycle, before methods for measuring bias in classical models and in LLMs are explored in depth. Further focal points include concrete mitigation strategies, legal and ethical frameworks (GDPR, EU AI Act, auditing standards), and current research trends such as mechanistic interpretability and AI governance. The course concludes with an independent mini-project on a self-selected use case. The course combines lecture videos with three practical exercises in Jupyter notebooks as well as weekly portfolio assignments, in which participants apply what they have learned and critically justify their own positions.

Gliederung:

Segment 0 – Introduction: organizational matters, technical requirements, examination arrangements, course structure. Segment 1 – Fundamentals of machine learning. Segment 2 – Introduction to bias and fairness in AI. Segment 3 – Causes of bias in AI. Segment 4 – Measuring bias in AI. Segment 5 – Measuring bias in LLMs. Segment 6 – Strategies for reducing bias in machine learning. Segment 7 – Strategies for reducing bias in LLMs. Segment 8 – Legal and ethical aspects. Segment 9 – Mini-project: case study in a self-selected field of application. Segment 10 – Future perspectives and current research trends (explainability/mechanistic interpretability, knowledge integration, fairness in modern LLMs, governance and participation). In addition: three practical exercises as Jupyter notebooks as well as weekly portfolio assignments, which together constitute the examination.

Detaillierter Inhalt:

Segment 0 – Introduction: project leadership, schedule, technical requirements, terminology, examination arrangements, grading criteria, accompanying literature, as well as an overview of the chapters and learning environment (videos, portfolio, forums, chatbot).
Segment 1 – Fundamentals of machine learning: historical overview of AI, ML, and deep learning through to foundation models and LLMs, using examples from vision, medicine, genomics, and NLP; types of learning, basics of neural networks, and the ethical challenges of modern LLMs as a transition to the course's focus.
Segment 2 – Introduction to bias and fairness in AI: conceptual clarification of bias (technical diagnosis) and fairness (normative judgment), key fairness definitions, and typical trade-offs (e.g., accuracy vs. group fairness). Sustainable fairness also requires institutional mechanisms.
Segment 3 – Causes of bias in AI: taxonomy of bias sources along the ML lifecycle (human/societal, data, model, deployment/feedback) and their application to LLMs (training data, tokenization, RLHF, sycophancy).
Segment 4 – Measuring bias in AI: fairness criteria (group, individual, utility-based, counterfactual fairness) using examples such as credit lending, the impossibility theorem, and real-world case studies (COMPAS, Obermeyer et al.).
Segment 5 – Measuring bias in LLMs: methods and benchmarks for measuring explicit and implicit bias, and their methodological limitations (sensitivity to phrasing, construct validity, predominantly English-language development).
Segment 6 – Strategies for reducing bias in machine learning: mitigation strategies along the ML pipeline (pre-, in-, post-processing) and common open-source toolkits (AIF360, Fairlearn, TensorFlow Responsible AI).
Segment 7 – Strategies for reducing bias in LLMs: prompting, retrieval-augmented generation, fine-tuning (LoRA/PEFT, RLHF), and agent-based approaches, plus relevant tools/benchmarks (HELM, TrustLLM, SafetyBench).
Segment 8 – Legal and ethical aspects: real-world cases of harm as a starting point for ethical analysis, legal frameworks (GDPR, EU AI Act), standards/auditing, and community-driven initiatives (NeurIPS-DEI, NLP4SG, FAccT).
Segment 9 – Mini-project (case study): independent project work in a freely chosen field of application (e.g., HR, healthcare, finance, facial recognition) with a complete workflow for detecting and mitigating bias; results submitted as a portfolio contribution.
Segment 10 – Future perspectives: mechanistic interpretability, knowledge integration (including GraphRAG), fairness in modern LLMs (long-term fairness, personalization, jailbreak testing), and governance and participation (including Collective Constitutional AI).

Lern-/Qualifikationsziele:

Upon successful completion of the course, participants will be able to: identify, classify, and explain different types of bias in artificial intelligence; analyze the causes and consequences of biased AI systems along the entire ML lifecycle; evaluate and apply methods for detecting and measuring bias in classical ML models as well as in large language models (LLMs); critically assess the ethical, legal, and societal implications of AI bias and incorporate them into their own development and evaluation concepts; design AI systems with fairness considerations in mind and implement mitigation strategies (pre-, in-, and post-processing); reflect on, critically position, and discuss differing, at times conflicting, conceptions of bias and fairness.

Lehrveranstaltungstyp:

Virtuelle Vorlesung

Interaktionsformen mit Betreuer/in:

Übungsaufgaben für Selbstlernbetrieb, Video-/Webkonferenz, E-Mail

Interaktionsformen mit Mitlernenden:

Forum

Kursdemo:

zur Kursdemo

Nutzung

Kurs ist konzipiert für:

Bachelor

Hochschule München: Data Science BA

TH Nürnberg: Informatik, Medieninformatik, Wirtschaftsinformatik: 6./7. Semester,
Fachwiss. Wahlpflichtfach

Formale Voraussetzungen:

Anmeldung im vhb-Kurs

Erforderliche Vorkenntnisse:

Prior knowledge of machine learning is required; the fundamentals of machine learning as well as large language models and foundation models are reviewed in Segment 1 of the course. Basic programming skills in Python are recommended, as the three practical exercises are carried out in Jupyter notebooks.

Hinweise zur Nutzung:

The course is structured into an orientation unit and ten content segments (approx. 2 hours of lecture, 4 hours of exercises, and 2–4 hours of portfolio work per week). The examination consists of a semester-long portfolio including three programming exercises in Jupyter notebooks; there is no separate final exam. A course-specific chatbot is available for questions about the course material. The course is offered bilingually (German/English).

Kursumsetzung (verwendete Medien):

Short, thematically self-contained lecture videos (screencasts with slides) in the Moodle instance of Munich University of Applied Sciences, supplemented by reading recommendations and discussion forums. After each lecture unit, participants complete a portfolio assignment, some with open reflection questions; in addition, there are practical exercises in Jupyter notebooks as well as an independent mini-project on a self-selected use case. A purpose-built chatbot, hosted on HM servers, supports independent work with the course material and, among other things, generates dynamic self-test quizzes. The videos were additionally recorded in German and automatically subtitled.

Erforderliche Technik:

Internet-enabled computer with an up-to-date web browser, access to Moodle. The practical exercises additionally require Anaconda with Jupyter Notebook (Python 3).

Nutzungsentgelte:

für andere Personen als (reguläre) Studenten der vhb Trägerhochschulen nach Maßgabe der Benutzungs- und Entgeltordnung der vhb

Rechte hinsichtlich des Kursmaterials:

-

Verantwortlich

Anbieterhochschule:

HAW München (HM)

Anbieter:

Prof. Dr. Anna Kruspe

Autoren:

Anna Kruspe

Betreuer:

M.Sc. Mila Stillman

Prüfung

Art der Prüfung:

praktischer Leistungsnachweis

Bemerkung:

Submission of portfolio until the given date via Moodle

Prüfer:

Prof. Dr. Anna Kruspe

Prüfungsanmeldung erforderlich:

ja

Anmeldeverfahren:

–

Prüfungsanmeldefrist:

–

Prüfungsabmeldefrist:

–

Kapazität:

–

Prüfungsdatum:

31.01.2027

Prüfungszeitraum:

–

Prüfungsdauer:

–

Prüfungsort:

–

Zuständiges Prüfungsamt:

–

Zugelassene Hilfsmittel:

everything, including course materials and chatbot

Formale Voraussetzungen für die Prüfungsteilnahme:

–

Inhaltliche Voraussetzungen für die Prüfungsteilnahme:

Background knowledge of machine learning basics and Python; knowledge of Large Language Model basics is desirable

Zertifikat:

Ja (proof of exam)

Anerkennung:

–