Konferenzen und Rahmenprogramm
Certifying Neural Networks for Autonomous Flight
Neural networks (NNs) have demonstrated impressive performance on visual recognition tasks relevant for autonomous drones or personal air-taxis. So, their application to visual problems like object detection and image segmentation is promising, and even necessary, for autonomous flight. Yet, for safety-critical applications, it is generally assumed to be difficult or impossible to demonstrate correctness and robustness of NNs. In this talk, we argue that the design and verification of NN-based systems need new methods to assure safety and performance. Over the last months, we have partnered with EASA (European Aviation Safety Agency) to explore concepts for design assurance and certification of NNs in airborne systems. We will present an overview of our framework for quantifying uncertainty associated with NNs and discuss model explainability, generalizability and adversarial attacks. Finally, we will explore how these ideas fit in with current certification processes such as DO-178C.
--- Datum: 25.02.2020 Uhrzeit: 16:30 - 17:00 Uhr Ort: Conference Counter NCC Ost