SHF: Design Automation of Resilient and Interpretable Neural Architectures for Safety-Critical Systems

Modern systems such as self-driving vehicles and automated medical devices increasingly rely on artificial intelligence to make important decisions. However, these systems can behave unpredictably when they encounter situations that were not seen during training, and their decision-making processes are often difficult for people to understand. This project addresses the need for artificial intelligence that is both reliable and understandable, particularly in applications where safety is critical. By improving the ability to detect and respond to unexpected conditions, this work aims to reduce failures in autonomous systems and enhance public trust in emerging technologies. The project also contributes to national priorities by improving transportation safety, advancing healthcare technologies, and preparing a workforce trained in trustworthy artificial intelligence. This project develops new methods for designing and evaluating neural network models that are resilient to faults and easier to interpret. The research focuses on a class of models in which each connection is represented by a learnable function, enabling direct inspection of how inputs influence outputs. A search framework is used to automatically generate model architectures that balance accuracy, robustness, and computational efficiency. The project integrates formal verification techniques to guarantee key safety properties and incorporates runtime monitoring to detect abnormal behavior during operation. A simulation-based testing platform is used to evaluate system performance under realistic conditions, including sensor noise, hardware faults, and rare scenarios. The methods are demonstrated in two application domains: autonomous driving and automated insulin delivery. The expected outcome is a set of tools and design principles for building artificial intelligence systems that operate safely, perform efficiently on edge hardware, and provide transparent decision-making. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2551665 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Fadi Kurdahi | Institution: University of California-Irvine, IRVINE, CA | Award Amount: $650,000 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2551665 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2551665.html