The intended audience is computer science or math students with background in algorithms, logic, and deep neural networks.
We assume familiarity with algorithms and logic. In-depth familiarity with deep learning is not required, but you should know how to train neural networks. Basic Python skills are required. Please contact the instructor should you be unsure if you have the necessary background.
The exceptional performance of deep neural networks in areas such as perception and natural language processing has made them an integral part of many real-world AI systems, including safety-critical ones such as medical diagnosis and autonomous driving. However, neural networks are inherently opaque, and numerous defects have been found in state-of-the-art networks.
In this lab, we will apply various methods for proving the reliability of deep neural networks. In particular, we will use state-of-the-art tools, such as Crown, ERAN, Marabou, and Planet, and apply them to examples from the neural network verification competition.