Risk-Aware Debugging Techniques for Deep Neural Networks

Défense de mémoire de Jonah Verly

In the autonomous driving field, vehicles relies on systems that are based on trained Deep Neural Networks (DNN). However autonomous driving field is a safety-critical domain, i.e. security is essential since human lives are at stake. In this master thesis, Distributed Repair of Deep Neural Networks and a Multi-Objective Optimization based repair technique are presented and compared. The comparison is made on the same data and the same model. However, some priorities and risks should be considered concerning the objects that these DNNs should detect. The comparison results shows that both techniques are similar performance but there are some points to take into account with some specific problems that are solved by the DNNs.

Contact : Isabelle Daelman - isabelle.daelman@unamur.be
Télecharger : vCal