Adresse :
Public cible
Amphitéatre du Pôle Numérique Brest Iroise (PNBI)
Reducing the complexity of Convolutional Neural Networks by quantization and filter pruning
Deep Neural Networks are powerful machine learning models that have obtained excellent performance for numerous applications. However, they may require a large amount of computational and memory resources leading to a high energy consumption making their use for embedded systems difficult. In this talk, I will present two different approaches that we proposed recently to reduce this complexity. This first is a post-training quantization method that approximates the filters of a CNN using dyadic rationals such that all multiplications of the model can be replaced by bit shifts and additions. The second approach is a compression-aware training algorithm, that performs filter pruning on a CNN by introducing a new sparsity-inducing regularization term based on the l1/l2 pseudo norm. Both approach have been applied to standard computer vision benchmarks and can considerably reduce the complexity of CNN with only a small precision loss.
3:50 pm: Eliott Coyac
A tour through the Hugging Face Hub
L'équipe BRAIn vous convie à un séminaire le 8 février 2023 au cours duquel nous avons le plaisir d'accueillir Stefan Duffner, Maître de Conférences à l'INSA Lyon au laboratoire LIRIS, ainsi que Eliott Coyac de Hugging Face.
Stefan nous parlera de compression de réseaux de neurones, tandis que Eliott nous présentera un nouvel outil de partage de modèle et datasets de Machine Learning développé par Hugging Face.
Organisateur(s)
Équipe de Recherche : BRAIn (Better Representations for Artificial Intelligence)
Partenaire(s)
INSA Lyon
LIRIS