MC 5501
Speaker
Xiangru Xu, University of Wisconsin-Madison
Title
Safety Verification of Neural Network Control Systems Using Constrained and Hybrid Zonotopes
Abstract
Neural networks are playing an increasingly prominent role in autonomous systems, particularly within perception and control stacks. However, ensuring the provable safety of neural network-enabled control systems remains a critical challenge. In this talk, I will present our recent work on computing forward and backward reachable sets for neural network control systems. These sets are computed either as a single hybrid zonotope or a finite union of constrained zonotopes. Furthermore, we have developed algorithms that enable a tunable trade-off between computational efficiency and approximation accuracy.