Safe Neural Networks for Autonomous Driving:
Supervisors
Suitable for
Abstract
Neural networks have proven to be very powerful at computer vision tasks. However, they often exhibit unexpected behaviours,
violating known requirements expressing background knowledge. This can be particular dangerous, especially in safety critical
scenarios such as autonomous driving. In this project, we will explore out to develop post-processing modules which can guarantee
that the final outputs are always compliant with the requirements and can be built on top of multi-label classification neural
networks . The post-processing modules will need to (i) be able to improve the performance of the model (in terms of the relevant
metric), and (ii) be able to run in linear time with respect to the number of labels. The proposed modules will be tested
on the ROad event Awareness Dataset with logical Requirements (ROAD-R), which is the first multi-label dataset for autonomous
driving with arbitrary n-ary constraints on the labels.