A minimal perturbation of inputs in safety critical AI can lead to catastrophic failures. In automotive systems, robustness aims to encompass reliable behavior under uncertainty, distributional shift, and rare edge cases.
How should robustness be defined, measured, and validated at a system-level in safety-critical AI? The session will explore how realistic errors can uncover failure modes in deep learning frameworks and how robustness measures can be applied to safety critical AI often under changing input conditions, highlighting the need for more rigorous, interpretable evaluation methods.
