Partager cette page :

How radiation can mess with your AI and what we can do about it!

le 6 novembre 2024

13h15

Campus de Beaulieu Salle Jersey - bât. 12D

Intervention de Fernando Fernandes dos Santos, chercheur Inria au centre Inria de l'Université de Rennes, dans le cadre des séminaires du département Informatique.

/medias/photo/seminaire-di_1630676501273-jpg

Deep Neural Networks (DNNs) have become state-of-the-art in AI, demonstrating high performance in various tasks, including language processing, multimodal analysis, image classification, and instance segmentation. Due to their high accuracy, DNNs are being increasingly employed in safety-critical applications such as space exploration, autonomous driving, and industrial automation, where high reliability and low power consumption are mandatory. Unfortunately, ionizing radiation can induce hardware faults that disrupt the internal operations of DNNs, leading to incorrect inferences. As traditional fault tolerance strategies based on modular hardware redundancy, software replications, or checksums impose unacceptable overheads, we need efficient and effective hardening solutions tailored for large DNNs. In this talk, Fernando Fernandes from the INRIA TARAN Team will discuss the challenges of evaluating and improving the reliability of large DNNs in safety-critical applications and space missions. He will cover innovative design choices and training procedures proposed by the TARAN team to improve DNNs' reliability with minimal overheads.
Thématique(s)
Formation, Recherche - Valorisation
Contact
David Pichardie

Mise à jour le 4 novembre 2024