Abstract – Traditional engineered systems are built in a modular, transparent fashion, each component, such as an airplane’s wing or wheel, has a clearly defined role that can be independently verified. In stark contrast, modern AI models are developed end-to-end through data-driven optimization, often resulting in highly effective but opaque systems whose internal logic is difficult to interpret or validate. This talk introduces SemanticLens, a novel method that brings engineering-style inspection to AI by projecting internal model representations into the semantically grounded space of foundation models such as CLIP. This projection allows us to analyze what knowledge a model encodes, how it leverages this knowledge during prediction, and where it originates in the training data. By enabling intuitive model comparison, debugging, and validation, SemanticLens supports deeper reasoning about model behavior and alignment with human expectations. We illustrate its utility with case studies in medical AI, showing how the method can uncover hidden failure modes, improve robustness, and ultimately narrow the gap between traditional engineering standards and AI system development.
Biography
Wojciech Samek is a Professor in the EECS Department at the Technical University of Berlin and the Head of the AI Department at the Fraunhofer Heinrich Hertz Institute (HHI) in Berlin, Germany. He earned an M.Sc. from Humboldt University of Berlin in 2010 and a Ph.D. (with honors) from the Technical University of Berlin in 2014. Following his doctorate, he founded the “Machine Learning” Group at Fraunhofer HHI, which became an independent department in 2021. He is a Fellow at BIFOLD – the Berlin Institute for the Foundation of Learning and Data and the ELLIS Unit Berlin. He also serves as a member of Germany’s Platform for AI and sits on the boards of AGH University’s AI Center, the Helmholtz Einstein School in Data Science (HEIBRiDS), and the DAAD Konrad Zuse School ELIZA. Dr. Samek’s research in explainable AI (XAI) spans method development, theory, and applications, with pioneering contributions such as Layer-wise Relevance Propagation (LRP), advancements in concept-level explainability, evaluation of explanations, and XAI-driven model and data improvement. He has served as a senior editor for IEEE TNNLS, held associate editor roles for various other journals, and acted as an area chair at NeurIPS, ICML, and NAACL. He has received several best paper awards, including from Pattern Recognition (2020), Digital Signal Processing (2022), and the IEEE Signal Processing Society (2025). Overall, he has co-authored more than 250 peer-reviewed journal and conference papers, with several recognized as ESI Hot Papers (top 0.1%) or Highly Cited Papers (top 1%).”
