About us
Espace utilisateur
INSTN offers more than 40 diplomas from operator level to post-graduate degree level. 30% of our students are international students.
Professionnal development
Professionnal development
Find a training course
INSTN delivers off-the-self or tailor-made training courses to support the operational excellence of your talents.
Human capital solutions
At INSTN, we are committed to providing our partners with the best human capital solutions to develop and deliver safe & sustainable projects.
Home   /   Thesis   /   Deep Neural Network Uncertainty Estimation on Embedded Targets

Deep Neural Network Uncertainty Estimation on Embedded Targets

Artificial intelligence & Data intelligence Computer science and software Engineering sciences Technological challenges


Over the last decade, Deep NeuralNetworks (DNNs) have become a popular choice to implement Learning-Enabled Components LECs in automated systems thanks to their effectiveness in processing complex sensory inputs, and their powerful representation learning that surpasses the performance of traditional methods. Despite the remarkable progress in representation learning, DNNs should also represent the confidence in their predictions to deploy them in safety-critical systems. Bayesian Neural Networks (BNNs) offer a principled framework to model and capture uncertainty in LECs. However, exact inference in BNNs is difficult to compute. Thus, we rely on sampling techniques to approximate the true posterior of the weights for computing the posterior predictive distribution (inference). In this regard, relatively simple though computationally and memory expensive sample-based methods have been pro posed for approximate Bayesian inference to quantify uncertainty in DNNs, e.g., Monte-Carlo dropout or Deep Ensembles. Efficient DNN uncertainty estimation in resource-constrained hardware platforms remains an open problem, limiting the adoption within applications from highly automated systems that possess strict computation and memory budgets, tight time constraints, and safety requirements. This thesis aims to develop novel methods and hardware optimizations for efficient and reliable uncertainty estimation in modern DNN architectures deployed in hardware platforms with limited computation resources.


Département Ingénierie Logiciels et Systèmes (LIST)
Labo.conception des systèmes embarqués et autonomes
Top envelopegraduation-hatlicensebookuserusersmap-markercalendar-fullbubblecrossmenuarrow-down