Sample-Efficient Uncertainty Calibration for Reliable Autonomous Systems (Rachel Luo, PhD Defense)

Rachel Luo PhD defense (08/28/2023) As the capabilities of autonomous systems continue to improve, they are increasingly deployed in unstructured environments for safety-critical tasks where the cost of failure is high. These systems take as input observations from the external environment, rely on intermediate learning-based components, and output actions that govern system behavior. To operate safely and reliably in high-stakes environments, these systems must have a meaningful and well-calibrated notion of uncertainty over the inputs, over the outputs, and for the internal components. In the first part of this talk, I will address the problem of calibrating the uncertainty representations used by intermediate components, enabling improved real-time decision-making. In the second part of this talk, I will address the problem of designing a meaningful and calibrated notion of uncertainty over the system outputs. Recognizing when the outputs of the system may lead to a dangerous mistake is a critical form of uncertainty; thus, I will present a warning system that issues alerts when a dangerous situation is imminent, and show that these alerts are calibrated by providing a statistical guarantee on their false negative rate. This calibration is achieved with limited data using techniques from conformal prediction. In the third part of this talk, I will address the problem of designing a meaningful and calibrated notion of uncertainty over the system inputs. Knowing when the underlying data distribution has changed and the system may be operating outside its domain of competency is a critical form of uncertainty, as such distribution shifts may lead to catastrophic failures for learning-based systems. To address this problem, I will present a warning system that issues alerts when the data distribution shifts, and show that these alerts are calibrated in the sense that there are rarely false alarms (backed with statistical guarantees). This alert system issues alerts an order of magnitude faster than prior work.