Séminaire R-D : Edge implementation of Deep Neural Network

Séminaire R-D : Edge implementation of Deep Neural Network
October 10, 2019 11:00
CRIM (405, avenue Ogilvy, bureau 101, Montréal)

EDGE IMPLEMENTATION OF DEEP NEURAL NETWORKS

 

Edge Implementation of Deep Neural Networks

Présentation en anglais

Réseaux de neurones profonds dans un contexte d'informatique en périphérie

 

 

CONFÉRENCIER

Vahid Partovi Nia, Ph. D, Scientifique principal en apprentissage automatique, Huawei


Edge implementation of Deep Neural Networks

 

 

SPEAKER

 

Dr. Vahid Partovi Nia, Principal Machine Learning Scientist, Huawei

ABSTRACT

 

This talk is divided into two parts. In the first part I focus on network quantization foundation. Low-bit quantization helps to speed up inference, save memory, and ultimately brings the intelligence on edge devices. Motivated from the perceptron, I argue that a quantized network is equivalent to a complex logic. This provides us with a powerful computation vehicle to deploy neural networks only using logical operators. In the second part of the talk I focus on designing compact neural network architectures. I study the exhaustive search foundation as the main powerful neural network design tool. Despite application of the exhaustive search algorithm in many complex problems there is little theory developed in this field. I show some basic theoretical results that give us an insight about convergence conditions of the exhaustive search algorithm.

BIOGRAPHY


Conférence gratuite. Inscription requise.

Upcoming event

  • Séminaire R-D : Fault detection in timed FSM with timeouts by sat-solving
    24 October 2019 11:00
    CRIM (405, avenue Ogilvy, bureau 101, Montréal)
    Une conférence de Omer Nguena Timo, chercheur en modélisation, test et vérification logiciel, CRIM.
    +
  • ???? Le CRIM est en mode recrutement #talents 2 nouveaux profils de stagiaires recherchés : - Science des données - Sy… https://t.co/JxKxg7ZwKw
  • ???? Séminaire R-D : Détection des défauts dans les FSM temporisés par résolution de contraintes SAT (présentation en… https://t.co/VIDB9R4u4u