Doctorant sous la direction de F. DUFAUX

Titre de la thèse: Study of deep learning methods applicable in the context of the autonomous vehicle's perception.
Résumé de la thèse: The objective of this thesis is to initiate research into deep learning methods applied to perception in the context of the autonomous vehicle. One of the major constraints of deep learning methods is the need of properly labeled data for the training phase of neural networks. In general, the data require to be fully labeled manually. Thus, my work is focused on the study of methods allowing to reduce as much as possible the need of hand-labeled data. It turns out that several methods already exist to answer this problem. The most popular include: - semi-supervised learning. This consists in pre-training the weights of the neural networks in an unsupervised way on unlabeled data. - self-supervised learning. This consists of learning about data automatically labeled by another method beforehand, which should avoid possible automatic labeling of false positives and false negatives so as not to disrupt training. This technique can be applied to perform incremental learning. Moreover, there are methods where the objective is to require only a part of the information, to be able to deduce the other information: - one-class learning. This consists of learning to distinguish the samples of the class of interest, the positive samples, from those of the counter-example class, the negative samples, by disposing only during the training phase of positive samples. - positive unlabeled learning. This consists of learning on a dataset composed of positive samples on the one hand, and unlabeled samples on the other hand, which can be positive or negative. - noisy labeled learning. Considered as weakly supervised learning, it consists in learning with a training dataset composed of a fraction of positive and negative samples which are badly labeled. Our work in progress consists of proposing new approaches that are transversal to those described in order to improve the prediction performance and robustness of a learning model in the context of applications on sensor data of the delegated vehicle.

Summary