Ringvorlesung 2026 "Künstliche Intelligenz - quo vadis?"
From Explainable to Trustworthy AI
Referent: Prof. Dr. Wojciech Samek (Fraunhofer HHI)
Kurzbeschreibung:
Traditional engineered systems, such as an airplane’s wing or a bridge’s foundation, are built from modular, transparent components, each with a well-defined and independently verifiable function. In contrast, modern AI models are developed end-to-end through data-driven optimization. While this often yields powerful capabilities, it also produces opaque systems whose internal logic is difficult to interpret or validate. This talk introduces new methods that bring engineering-style inspection to AI models, enabling a deeper understanding of their internal representations and behaviors. In large language models, for example, we can identify specialized attention heads: in-context heads, which interpret instructions and retrieve relevant contextual information through retrieval-augmentation, and parametric heads, which encode relational knowledge about entities. This fine-grained insight not only illuminates how LLMs reason but also supports practical tools for detecting and mitigating hallucinations, advancing the development of safer and more trustworthy AI.