The challenge of building stable, accurate and robust data-driven AI

In this talk, we will delve into the theoretical limitations of determining the guaranteed stability and accuracy of neural networks built from empirical data in classification tasks. We will show that there is a large family of tasks and settings in which computing and verifying stability and accuracy is extremely challenging. We will also discuss an intriguing connection of these results with adversarial data and examples and propose a potential way to remedy the issues by enabling the networks to adapt over time.