Energy-Efficient, Flexible and Fast Architectures for Deep Convolutional Neural Network Acceleration
Year: 2020
Type: dissertation
Abstract: RESUME: Les methodes basees sur l'apprentissage profond, et en particulier les reseaux de neurones convolutifs (CNN), ont revolutionne le domaine de la vision par ordinateur. Alors que jusqu'en 2012, les methodes de traitement d'image traditionnelles les plus precises pouvaient atteindre 26% d'erreurs dans la reconnaissance d'images sur l'etalon normalise et bien connu ImageNet, une methode basee sur un CNN a considerablement reduit l'erreur a 16%. En faisant evoluer la structure des CNN, les methodes actuelles basees sur des CNN atteignent desormais couramment des taux d'erreur inferieurs a 3%, depassant souvent la precision humaine. Les CNN se composent de nombreuses couches convolutives, chacune effectuant des operations de convolution complexes de haute dimension. Pour obtenir une precision elevee en reconnaissance d’images, les CNN modernes empilent de nombreuses couches convolutives, ce qui augmente considerablement la diversite des motifs de calcul entre les couches. Ce haut niveau de complexite dans les CNN implique un nombre massif de parametres et de calculs.----------ABSTRACT: Deep learning-based methods, and specifically Convolutional Neural Networks (CNNs), have revolutionized the field of computer vision. While until 2012, the most accurate traditional image processing methods could reach 26% errors in recognizing images on the standardized and well-known ImageNet benchmark, a CNN-based method dramatically reduced the error to 16%. By evolving CNNs structures, current CNN-based methods now routinely achieve error rates below 3%, often outperforming human level accuracy. CNNs consist of many convolutional layers each performing high dimensional complex convolution operations. To achieve high image recognition accuracy, modern CNNs stack many convolutional layers which dramatically increases computation pattern diversity across layers. This high level of complexity in CNNs implies massive numbers of parameters and computations. Since mobile processors are not designed to perform massive computations, deploying CNNs on portable and mobile devices is challenging.
Author: Mehdi Ahmadi
Institution: Polytechnique Montréal
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3159510977
