Safe and Trustworthy AI with Verifiable Guarantees
 

Abstract

In this talk, I will present my work on developing safe and trustworthy AI systems with verifiable guarantees, situated at the intersection of AI and formal methods. I’ll begin with a brief overview of my research scope, then highlight two key areas: (1) formal explainable AI to promote trustworthiness, and (2) robustness guarantees to enhance AI safety. In the first part, I’ll discuss how we use neural network verification techniques to compute optimal verified explanations for deep neural networks, and how these explanations can be applied to evaluate trustworthiness in real-world scenarios—such as an autonomous aircraft taxiing application developed in collaboration with Stanford AeroAstro and NASA. In the second part, I’ll introduce a game-based verification framework that computes the maximum safe radius of neural networks to quantify their robustness. Notably, the concept of "safe radius" has been adopted by ISO and IEC in their newly established standard ISO/IEC TR 5469:2024 Artificial Intelligence – Functional Safety and AI Systems. I will conclude with a discussion of future research directions.
 

Biography

Min Wu, postdoctoral scholar, works with Professor Clark Barrett in the Department of Computer Science at Stanford University. She is also affiliated with the Stanford Center for AI Safety and the Stanford Center for Automated Reasoning. She received her PhD in computer science from the University of Oxford under the supervision of Professor Marta Kwiatkowska. Her research aims to develop AI systems—particularly those used in high-stakes applications—that are verifiably safe and trustworthy. More details about her work are available on her academic webpage.