While discriminative networks are a huge success (e.g. in image classification, segmentation, object detection), they have certain shortcomings, e.g., the confidence score is not calibrated, out-of-distribution (OOD) detection is not solved.

Recent publications, e.g., "Your Classifier is Secretly an Energy Based Model and You Should Treat it Like One.; ICLR 2020 by Grathwohl et al." show that training hybrid (generative and discriminative) networks can have benefits, e.g. better calibrated confidences, OOD detection, being more robust to adversarial attacks.

Inspired by this, the student should investigate such methods in the domain of computer vision applied for self-driving, e.g. do OOD for object detectors, semantic segmentation.

Solid background in mathematics and python programming is required.