Risk-Driven Design of Perception Systems
2022: Advances in Neural Information Processing Systems (NeurIPS)
We developed a method to design safer perception systems using a risk-driven approach that accounts for closed-loop safety properties. We applied our methods to a vision-based detect and avoid application and showed a 37% decrease in collision risk over a baseline model.
Verification of Image-Based Neural Network Controllers Using Generative Models
2022: Journal of Aerospace Information Systems
2021: Digital Avionics Systems Conference (DASC)
We created a method to formally verify image-based neural network controllers by using a generative model to capture the set of plausible input images. We applied our methods to provide guarantees on a vision-based taxi navigation system.
Generating Probabilistic Safety Guarantees for Neural Network Controllers
2021: Machine Learning Journal
We developed a technique to analyze the safety of neural network controllers used in stochastic environments. We applied the technique to analyze the safety of a neural network-based aircraft collision
avoidance system.
Sydney M. Katz 