Titre:
On the Efficiency of Filter Pruning in Deep Convolutional Neural Networks
Conférencier:
Vincent Gripon ,
IMT Atlantique
Lieu:
Polytechnique Montréal, Pavillon J.-Armand Bombardier, local 1035 ,
Date et heure:
vendredi le 26 mai 2023 de
10:30 à 12:00
Résumé: Pruning is a cornerstone of neural network compression. When it is performed in a structured way, it is expected to generate lighter and more efficient architectures, with almost no impact on performance. Yet, very few papers demonstrate this ability to impact physical measures of chips, including power consumption and latency. In this presentation, we shall revisit this scientific challenge. After a quick overview about pruning, we will introduce a novel methodology, explain how to leverage the resulting architectures on generic hardware, and discuss results.
Note biographique: Vincent Gripon is a full professor with IMT Atlantique, one of the best French technological university. He obtained his Master of Science with École Normale Supérieure Paris-Saclay in 2008 and his PhD with Télécom Bretagne in 2011. He spent one year as a postdoctoral researcher in McGill University in 2012. He also has been an invited professor at Université de Montréal and Mila in 2019. His research interests are at the crossbreed of Artificial Intelligence, Neural Networks, Few-Shot Learning and Efficient Implementations.