I held a presentation at the 2. Deep Learning Day of Hochschule der Medien on Machine Learning for Autonomous Driving, giving an overview on promising ML-based methods and some recent approaches that have been developed at the Bosch Center for Artificial Intelligence. (Hochschule der Medien, Stuttgart, Germany, January 12, 2018)
Machine Learning for Autonomous Driving (Michael Herman, Dr. Bastian Bischoff, Robert Bosch GmbH, Differentiating AI – Environmental Understanding – Decision Making) – For bringing autonomous vehicles on public roads, many questions need to be answered. For example, what are essential features in the high-dimensional observations from varying sensors, what are other traffic participants going to do, how are they going to react to the behavior of the autonomous car, how should an autonomous system behave, or how to find optimal strategies according to these criteria, while guaranteeing safety requirements. Machine Learning approaches can be used to answer these questions to some extent. While Deep Neural Networks have outperformed traditional methods in various applications, they have also shown to be vulnerable against adversarial perturbations on problems with high-dimensional input spaces. Since this can prevent its usage in safety- and security-critical applications such as autonomous driving, it is necessary to both understand the limits of these models and to increase their robustness.