Jun Liu
Professor of Applied Mathematics
Room: MC 5501
Can we verify AI?
Artificial intelligence (AI) is increasingly used to make decisions in safety-critical applications, where even a rare failure can have serious consequences. Yet even extensive testing cannot guarantee reliable behaviour in every case. In this talk, I will introduce formal verification of neural networks as a way to establish whether an AI system satisfies precisely stated requirements within a specified operating region. Using examples from neural network robustness and the stability and control of dynamical systems, I will explain how learning can generate candidate mathematical certificates and how formal verification can establish rigorous guarantees. I will then turn to a related question: can we verify mathematics produced by AI? I will share my experiences using AI to tackle long-standing open problems in dynamical systems and control, and invite an open discussion about what these developments might mean for mathematical research and education. No background in AI, control theory, or formal verification is assumed.