TITLE: Certified Artificial Intelligence
Despite remarkable success over the last decade, increasing evidence pointing to the fragility of modern data-driven AI systems has started to emerge, triggering social concerns, government regulations, and limiting wider adoption. Indeed, creating AI systems that behave safely and reliably is a fundamental challenge of critical importance. In this talk, I will present a path towards addressing this fundamental problem. Specifically, I will introduce new mathematical methods based on convex relaxations, sampling, and Lipschitz optimization that enable scalable and precise reasoning about the (potentially infinite number of) behaviors of an AI system (e.g., a deep neural network). I will then show how these methods enable both the creation of state-of-the-art automated verifiers for modern AI systems and the discovery of new provable training techniques. Finally, I will outline several promising future research directions.
Gagandeep Singh is a Ph.D. candidate at ETH Zurich supervised by Professors Martin Vechev and Markus Püschel. He completed his master’s from ETH and bachelor’s from the Indian Institute of Technology, Patna. His research interests lie at the intersection of automated reasoning and machine learning. He has built several systems now used in both academia and industry, including ELINA, a state-of-the-art library for fast numerical static analysis and ERAN, a state-of-the-art verifier for deep neural networks.