Skip to content

Exploring and Visualizing Referring Expression Comprehension (Bachelor's Thesis by David Álvarez Rosa)

Notifications You must be signed in to change notification settings

david-alvarez-rosa/referring-expression-comprehension

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Exploring and Visualizing Referring Expression Comprehension

This repository contains the source code develop in my senior Bachelor’s Thesis at University of Toronto entitled: “Exploring and Visualizing Referring Expression Comprehension”.

Abstract

Human-machine interaction is one of the main objectives currently in the field of Artificial Intelligence. This work will contribute to enhance this interaction by exploring the new task of Referring Expression Comprehension (REC), consisting of: given a referring expression—which can be a linguistic phrase or human speech—and an image, detect the object to which the expression refers (i.e., achieve a binary segmentation of the referred object). The multimodal nature of this task will require the use of different deep learning architectures, among them: convolutional neural networks (computer vision); and recurrent neural networks and the Transformer model (natural language processing).

This thesis is presented as a self-contained document that can be understood by a reader with no prior knowledge of machine learning. The bulk of the work consists of an exhaustive study of the REC task: from the applications; until the study, comparison and implementation of models; going through a complete description of the current state of the art. Likewise, a functional, free and public web page is presented in which interaction is allowed in a simple way with the model described in this work.

Resumen

La interacción humano-máquina es uno de los objetivos principales actualmente en el ámbito de la Inteligencia Artifcial. En este trabajo se contribuirá a facilitar esta interacción explorando la novedosa tarea de Comprensión de la Expresión Referente (CER), consistente en: dada una expresión referente —que puede ser una frase lingüística o habla humana— y una imagen, detectar el objeto al que la expresión se refiere (i.e., conseguir una segmentación binaria del objeto referido). El caracter multimodal de este cometido hará necesario el uso de diferentes arquitecturas de aprendizaje profundo, entre ellas: redes neuronales convolucionales (visión artificial); y redes neuronales recurrentes y el modelo Transformer (procesamiento del lenguaje natural).

Esta tesis se presenta como un documento autosuficiente que puede ser entendido por un lector sin conocimientos previos en aprendizaje automático. El grueso del trabajo consiste en un estudio exhaustivo de la tarea de CER: desde las aplicaciones; hasta el estudio, comparación e implementación de modelos; pasando por una descripción completa del estado del arte actual. Así mismo, se presenta una página web funcional, gratuita y pública en la que se permite la interacción de una manera sencilla con el modelo descrito en este trabajo.

Resum

La interacció humà-màquina és un dels objectius principals actualment en l’àmbit de la Intel.ligència Artifcial. En aquest treball es contribuirà a facilitar aquesta interacció explorant la nova tasca de Comprensió de l’Expressió Referent (CER), que consisteix en: donada una expressió referent —que pot ser una frase lingüística o parla humana— i una imatge, detectar l’objecte a què l’expressió es refereix (i.e., aconseguir una segmentació binària de l’objecte referit). El caràcter multimodal d’aquesta comesa farà necessari l’ús de diferents arquitectures d’aprenentatge profund, entre elles: xarxes neuronals convolucionals (visió artificial); i xarxes neuronals recurrents i el model Transformer (processament de el llenguatge natural).

Aquesta tesi es presenta com un document autosuficient que pot ser entès per un lector sense coneixements previs en aprenentatge automàtic. El gruix de la feina consisteix en un estudi exhaustiu de la tasca de CER: des de les aplicacions; fins a l’estudi, comparació i implementació de models; passant per una descripció completa de l’estat de l’art actual. Així mateix, es presenta una pàgina web funcional, gratuïta i pública en la qual es permet la interacció d’una manera senzilla amb el model descrit en aquest treball.

Acknowledgements

I would like to express my gratitude to Prof. Sanja Fidler for giving me the opportunity to carry out this project under her supervision and allowing me to be part of her laboratory and connect with its members. I also want to express my thanks to the entire Vector Institute staff for allowing me to use their computational resources, as well as for remotely assisting me with any technical problems that arose.

Likewise, I want to thank Prof. Xavier Giró for his work as liaison co-supervisor between Canada and Barcelona and for being part of the evaluation panel of this thesis.

Moreover, I want to express my gratitude to Fundació Privada Cellex and the Interdisciplinary Higher Education Centre. They have been the engines of my academic education and those that have allowed me to be part of this adventure of studying two official bachelors’ degrees simultaneously in the Politechnical University of Catalonia. Within this great team I want to give special thanks to Toni Pascual for his management of the mobility stay and the complications arising from the COVID-19 pandemic. Thanks also to Miguel Ángel Barja—with whom I have been lucky to be his student—for his role as director of the center.

Finally, I want to thank my family and friends for their unconditional moral support without asking me too much “When will you graduate?”—or at least not very often.

Copyright

David Álvarez Rosa © May 16, 2021 Exploring and Visualizing Referring Expression Comprehension https://recomprehension.com

Thesis typeset with pdfTeX 3.14159265–2.6–1.40.21 (TeX Live 2020) on Arch Linux using Latin Modern typefaces and written with GNU Emacs. The BibLATeX package has been used for bibliography management with Biber as processing backend.

Vector graphics have been created by the author using PGF/TikZ. Vectorian decorative ornaments are from the LaTeX package pgfornament.

This thesis is licensed under a Creative Commons “Attribution–NonCommercial–ShareAlike 4.0 International” license.