AMATH Prof Talk: Can We Verify AI?

Hello everyone! Our AMATH Prof Talk with Professor Jun Liu is coming up soon! The talk is on "Can we Verify AI?". Abstract: Modern AI systems can make impressive predictions, yet they often behave as black boxes: we can test them, but we usually cannot guarantee what they will do in every situation. This raises a natural question: can we verify an AI system, the way we verify a mathematical proof or a computer program? In this talk, we introduce the basic ideas behind neural network verification. Rather than relying only on testing or statistics, verification aims to prove that a trained model always satisfies certain properties, such as safety, robustness to perturbations, or consistency with physical laws. We explain how such guarantees are possible despite the complexity of neural networks, and why this is fundamentally different from standard machine learning evaluation. To make the ideas concrete, we use examples from control of dynamical systems, where incorrect decisions can cause instability and unsafe behavior. We show how verification can certify that a learned controller stabilizes a system or keeps it within safe operating limits. The goal of the talk is to give an intuitive picture of how mathematics can turn AI from a tool we hope works into a system we can trust. Food will be provided to members. Bring your friends and learn some math! Hope to see you there! Date: March 25 Time: 6:00 - 7:00 PM Location: MC 2017