At the 10th International Summer School on AI and Big Data, Adrian Lindenmeyer and Dr. Daniel Schneider will talk about Utility of knowledge uncertainty estimation in human-computer collaborative decision-making.

Talk: Utility of knowledge uncertainty estimation in human-computer collaborative decision-making

In the context of partially or supervised automated information processing, the aim is to automate as much routine work as possible while reserving review or resolution of ambiguous cases for manual intervention. In the application of AI-driven systems, this context emerges whenever stakes are high and consequences of mistakes could potentially be severe, such as in the fields of clinical decision support or autonomous driving. For effective collaboration, a fundamental challenge in such settings is enabling AI-driven agents to recognize and communicate the boundaries of their knowledge. Here, traditional point estimation methods fall short, commonly displaying overconfidence under data shift (cf. hallucination in AI). This calls for AI-driven solutions with means to estimate knowledge uncertainty and raises the question: How can we assess the utility of AI models within these human-AI collaborative frameworks and effectively compare various approaches?

This session addresses the pitfalls in human-computer interactive decision-making caused by overconfident AI models and demonstrates how adopting knowledge uncertainty-aware strategies can mitigate these issues. We explore methodologies to evaluate the practical utility of such strategies in semi-automated environments, providing insights into benchmarking techniques for comparative analysis. This discussion aims to shed light on improving the trustworthiness of AI agents and the effectiveness of human-computer collaboration in safety-critical applications.

Adrian Lindenmeyer

Adrian Lindenmeyer studied Engineering Science and Mechanical Engineering at the Technical University of Munich. As a Visiting Researcher at the Northwestern University in Chicago Illinois he wrote his Master’s Thesis on the topic of automated image evaluation using Bayesian reasoning gaining a Master of Science in 2020. He subsequently joined the MPM group at ICCAS in late 2020. His areas of focus are AI-based medical decision support utilising time-aware patient data (EHR). A major part of his work considers enabling Neural Networks to express and quantify uncertainty for a given prediction in an effort to increase the trustworthiness and safety of Neural Networks for medical use.

RESEARCH AREAS
– Statistical Modelling
– Artifical Intelligence
– Scientific Computing
– Uncertainty Quantification

Dr. Daniel Schneider

Daniel Schneider studied at Leipzig University, University of Amherst, MA, and University of Otago, NZ. He finished his PhD in Physics in November 2018 and subsequently turned his attention to medical image processing. In May 2020 Daniel Schneider joined the Models for Personalized Medicine (MPM) group at ICCAS as postdoc.
Dr. Schneider’s area of expertise encompasses statistical modeling and simulations, scientific computing, as well as machine learning and neural networks.