Titre:
Designing Robust DNN Models That Exploit Energy-Reliability Tradeoffs
Conférencier:
François Leduc-Primeau ,
Polytechnique Montréal
Lieu:
Polytechnique Montréal, Pav. Principal A-410 & Zoom ,
Date et heure:
mardi le 26 septembre 2023 de
10:30 à 11:30
Résumé:
Visionner l’enregistrement
Current state-of-the-art artificial deep neural networks (DNNs) tend to be energy intensive even at inference time, while they usually do not surpass the capabilities of the human brain, which functions with about 20 watts. To bridge this power gap, it is crucial to improve the energy efficiency of the computing systems that run the DNNs. Many new computing approaches have been proposed such as near-threshold CMOS circuits or processing-in-memory architectures, but these have in common that they are affected by different forms of noise. In this talk, I will review our recent proposals for analyzing and improving the robustness of DNN models deployed in these noisy computing systems.
Note biographique: François Leduc-Primeau was an Assistant Professor from 2019 to 2023 and is now an Associate Professor at École Polytechnique de Montréal. He received the B.Eng., M.Eng., and Ph.D. degrees in electrical & computer engineering from McGill University, Montreal, Qc, Canada, respectively in 2007, 2010, and 2016. From 2016 to 2018, he was a postdoctoral researcher at IMT Atlantique in Brest, France, and then at École de technologie supérieure in Montreal. His research interests span digital system design, telecommunications, and machine learning, with applications in next-generation wireless communication systems, energy-efficient artificial intelligence, and other low-energy digital systems. He is an IVADO professor, a member of ReSMiQ and OIQ, and has organized many research gatherings, such as the 2021 International Symposium on Topics in Coding (ISTC), 2022 ISTC Workshop, and the Hardware-Aware Efficient Training workshops at ICLR 2021 and ICML 2022.