Inspect, Understand, Overcome: A Survey of Practical Methods for AI Safety

Abstract

Deployment of modern data-driven machine learning methods, most often realized by deep neural networks (DNNs), in safety-critical applications such as health care, industrial plant control, or autonomous driving is highly challenging due to numerous model-inherent shortcomings. These shortcomings are diverse and range from a lack of generalization over insufficient interpretability and implausible predictions to directed attacks by means of malicious inputs. Cyber-physical systems employing DNNs are therefore likely to suffer from so-called safety concerns, properties that preclude their deployment as no argument or experimental setup can help to assess the remaining risk. In recent years, an abundance of state-of-the-art techniques aiming to address these safety concerns has emerged. This chapter provides a structured and broad overview of them. We first identify categories of insufficiencies to then describe research activities aiming at their detection, quantification, or mitigation. Our work addresses machine learning experts and safety engineers alike: The former ones might profit from the broad range of machine learning topics covered and discussions on limitations of recent methods. The latter ones might gain insights into the specifics of modern machine learning methods. We hope that this contribution fuels discussions on desiderata for machine learning systems and strategies on how to help to advance existing approaches accordingly.

A review of product safety regulations in the European Union

“…But according to the Commission's auxiliary report [23], the safety provisions for artificial intelligence applications, robotics, and Internet-of-things devices are seen as attainable mainly through the transparency, accountability, and unbiasedness of algorithms, fallback mechanisms, and keeping a human in the loop. Although these are sensible goals as such, it remains disputed how these could be legislated with sufficient K rigor-especially when considering that safety science is not quite there yet [33]. For these and other reasons, liability for new technologies has been actively debated recently.…”

Yes we care!-Certification for machine learning methods through the care label framework

“…The importance of trustworthy AI methods is increasing, especially because decision-making takes data-based models more and more into account (Bellotti and Edwards, 2001 ; Floridi et al, 2018 ; Lepri et al, 2018 ; Houben et al, 2021 ). In her comprehensive book, Virginia Dignum addresses AI's ethical implications of interest to researchers, technologists, and policymakers (Dignum, 2019 ).…”

FisheyePixPro: Self-supervised pretraining using Fisheye images for semantic segmentation

“…The domain of autonomous driving in- volves object detection [17,18,19], soiling detection [20,21,22], semantic segmentation [23], weather classification [24,25], dynamic object detection [26], depth prediction [27,28,29,30,31], fusion [32], key-point detection and description [33] and multitask learning [34,35,36]. It also poses many challenges due to the highly dynamic and interactive nature of surrounding objects in the automotive scenarios [37].…”