Coursera Spring Sale
40% Off Coursera Plus Annual!
Grab it
Explore the critical challenge of understanding and interpreting modern AI systems in this 29-minute keynote conference talk that addresses one of the most pressing issues in today's AI landscape: the lack of transparency in large language models and deep learning systems. Discover how Professor Wojciech Samek from TU Berlin presents groundbreaking engineering-style inspection methods that bring unprecedented clarity to AI model behavior, moving beyond traditional black-box approaches to reveal internal mechanisms at the component level. Learn about cutting-edge explainable AI (XAI) techniques that enable you to understand how classifiers use trusted versus spurious features, identify functional groups of neurons and attention heads, and distinguish between parametric and in-context reasoning in LLMs. Gain insights into detecting and mitigating AI hallucinations through citation-based explanations, using foundation models like CLIP to automatically label concepts learned inside neural networks, and building practical tools for debugging, correcting, and validating model behavior. Examine real-world case studies including dermatology AI systems and their hidden spurious correlations, while exploring the evolution from first-wave to third-wave XAI methodologies that enable fully automated concept discovery and model understanding. Master techniques for source tracing in LLMs, understanding how these models hallucinate and methods to control or prevent such behavior, and implementing engineering-level transparency comparable to traditional disciplines like aviation and bridge design. This comprehensive presentation showcases research from top-tier venues including NeurIPS and Nature Machine Intelligence, making it essential viewing for professionals working in large language models, responsible AI, model interpretability, AI safety and evaluation, medical AI validation, and anyone seeking to build more trustworthy and transparent AI systems.
