The revolutionary capabilities of AI with machine learning have enabled an increasingly broad range of applications, which has brought many new challenges in ensuring the trustworthiness of AI applications. In this talk, I will present our research on trustworthy AI with verifiable guarantees. I will first introduce our frameworks for the automatic formal verification of AI models as general computational graphs, to support general neural network architectures, nonlinearities, and safety properties being verified. I will also talk about our work on testing the soundness of neural network verifiers for the reliability of verifiers themselves. Then, I will present our verification-aware neural network training techniques for producing verification-friendly AI models with stronger verifiability. Finally, I will also discuss applications of our verification and verification-aware training in synthesizing verifiably stable neural network-based controllers for nonlinear dynamical systems.
Zhouxing Shi recently joined UC Riverside as an Assistant Professor in Computer Science and Engineering in July 2025. He completed his Ph.D. at the UCLA Computer Science Department. His research focuses on machine learning and trustworthy AI for building more reliable AI models. His recent research topics mostly involve the robustness, safety, and verification for AI models.