Optimization of Convolutional Neural Networks for - Brest, France - Lab-STICC, Université de Bretange Occidentale

Description

Optimization of convolutional neural networks for embedded platforms:

• Réf
  ABG-121115
• Stage master 2 / Ingénieur
• Durée 6 mois
• Salaire net mensuel standard French internship support 12/03/2024
• Lab-STICC, Université de Bretange Occidentale
• Lieu de travail
• Brest Bretagne France
• Champs scientifiques
• Informatique 30/03/2024
  Établissement recruteur:

Site web:

Le Lab-STICC, fort de son rattachement à l'institut INS2I du CNRS, est une unité de recherche historiquement reconnue en Bretagne Océane et en France dans le domaine des STIC.

Elle affiche une capacité avérée de couvrir un large spectre scientifique autour des sciences du numérique, et avec en particulier cette faculté d'adresser des champs disciplinaires variés (Théorie de l'Information, Ondes & Matériaux, Electronique et Informatique embarquées, Sciences des données, Communication et détection de signaux, Interfaces Homme-Machines,..)

suivant des thématiques/secteurs applicatifs multiples :
l'environnement maritime, les objets communicants, la défense, le spatial, la santé, la sécurité, la robotique...


Description:


Keyword:
convolutional neural network, image segmentation, embedded systems

This internship aims to optimize the execution time of convolutional neural networks on embedded platforms.

U-Net, a neural network shaped like the letter U developed for the segmentation of biomedical images [1], is now considered the reference (baseline) in thousands of articles in different fields in deep learning such as computer vision, signal and image processing.

The internship will begin with a state of the art on the following aspects:

• U-NET convolutional neural network and compression techniques (pruning, quantification, and knowledge distillation)
• Platform for the training phase (Anaconda, Python, Notebook, Tensorflow [2] or Pytorch). The training will be done on PIAF a dedicated GPU server for IA and machine learning of Lab-STICC, UBO.
• Experiments with embedded platforms. The targeted platforms are Beaglebone Black and Raspberry Pi.

After this study, we plan to achieve the following objectives

• Deployment of optimized CNN on the targeted embedded platforms
• Measure and compare the execution times and qualities of different optimization techniques including compressing and pruning
• Identify the key parameters in the optimization of the CNN on targeted embedded platforms
• Modification of the U-NET architecture to allow the production of intermediate results

Duration:
between 5 and 6 months


Support:
standard internship support following the French regulation for public establishments (4.35€/h)


Host institution:
Lab-STICC, Université de Bretagne Occidentale, Brest, France


Reference
[1] Olaf Ronneberger, Philipp Fischer et Thomas Brox, « U-Net: Convolutional Networks for Biomedical Image Segmentation », Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, Springer International Publishing,‎ 2015, p


Profil:


Prise de fonction:

• Dès que possible