Show item record

dc.contributor.advisorBengio, Yoshua
dc.contributor.advisorCourville, Aaron
dc.contributor.authorXu, Kelvin
dc.date.accessioned2018-05-31T13:29:01Z
dc.date.availableNO_RESTRICTIONfr
dc.date.available2018-05-31T13:29:01Z
dc.date.issued2018-03-21
dc.date.submitted2017-12
dc.identifier.urihttp://hdl.handle.net/1866/20194
dc.subjectReseaux de Neuronesfr
dc.subjectGeneration de Descriptionfr
dc.subjectApprentissage Profondfr
dc.subjectApprentissage de Representationsfr
dc.subjectApprentissage Supervisefr
dc.subjectInference Variationellefr
dc.subjectApprentissage par Renforcementfr
dc.subjectAttentionfr
dc.subjectModelisation de Donnees Sequentiellesfr
dc.subjectNeural Networksfr
dc.subjectCaption Generationfr
dc.subjectDeep Learningfr
dc.subjectRepresentation Learningfr
dc.subjectSupervised Learningfr
dc.subjectVariational Inferencefr
dc.subjectReinforcement Learningfr
dc.subjectAttentionfr
dc.subjectSequence Modellingfr
dc.subject.otherApplied Sciences - Artificial Intelligence / Sciences appliqués et technologie - Intelligence artificielle (UMI : 0800)fr
dc.titleExploring Attention Based Model for Captioning Imagesfr
dc.typeThèse ou mémoire / Thesis or Dissertation
etd.degree.disciplineInformatiquefr
etd.degree.grantorUniversité de Montréalfr
etd.degree.levelMaîtrise / Master'sfr
etd.degree.nameM. Sc.fr
dcterms.abstractComprendre ce qu’il y a dans une image est l’enjeu primaire de la vision par ordinateur. Depuis 2012, les réseaux de neurones se sont imposés comme le modèle de facto pour de nombreuses applications d’apprentissage automatique. Inspirés par les récents travaux en traduction automatique et en détection d’objet, cette thèse s’intéresse aux modèles capables de décrire le contenu d’une image et explore comment la notion d’attention peut être parametrisée par des réseaux de neurones et utilisée pour la description d’image. Cette thèse presente un reseau de neurones base sur l’attention qui peut décrire le contenu d’images, et explique comment apprendre ce modèle de facon déterministique par backpropagation ou de facon stochastique avec de l’inférence variationnelle ou de l’apprentissage par renforcement. Etonnamment, nous montrons que le modèle apprend automatiquement a concentrer son attention sur les objets correspondant aux mots dans la phrase prédite. Cette notion d’attention obtient l’état de l’art sur trois benchmarks: Flickr9k, Flickr30k and MS COCO.fr
dcterms.abstractUnderstanding the content of images is arguably the primary goal of computer vision. Beyond merely saying what is in an image, one test of a system's understanding of an image is its ability to describe the contents of an image in natural language (a task we will refer to in this thesis as \image captioning"). Since 2012, neural networks have exploded as the defacto modelling tool for many important applications in machine learning. Inspired by recent work in machine translation and object detection, this thesis explores such models that can describe the content of images. In addition, it explores how the notion of \attention" can be both parameterized by neural networks and usefully employed for image captioning. More technically, this thesis presents a single attention based neural network that can describe images. It describes how to train such models in a purely deterministic manner using standard backpropagation and stochastically by considering techniques used in variational inference and reinforcement learning. Surprisingly, we show through visualization how the model is able to automatically learn an intuitive gaze of salient objects corresponding to words in the output sequence. We validate the use of an attention based approach with state-of-the-art performance three benchmark datasets: Flickr9k, Flickr30k and MS COCO.fr
dcterms.languageengfr


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show item record

This document disseminated on Papyrus is the exclusive property of the copyright holders and is protected by the Copyright Act (R.S.C. 1985, c. C-42). It may be used for fair dealing and non-commercial purposes, for private study or research, criticism and review as provided by law. For any other use, written authorization from the copyright holders is required.