Dissertations / Theses on the topic 'Visual question answering (VQA)'

This thesis explores a computer's ability to understand multimodal data where the correspondence between image/video content and natural language text are utilised to answer open-ended natural language questions through question-answering tasks. Static image data consisting of both indoor and outdoor scenes, where complex textual questions are arbitrarily posed to a machine to generate correct answers, was examined. Dynamic videos consisting of both single-camera and multi-camera settings for the exploration of more challenging and unconstrained question-answering tasks were also considered. In exploring these challenges, new deep learning processes were developed to improve a computer's ability to understand and consider multimodal data.

Alors que nous nous représentons le monde au travers de nos sens, de notre langage et de nos interactions, chacun de ces domaines a été historiquement étudié de manière indépendante en apprentissage automatique. Heureusement, ce cloisonnement tend à se défaire grâce aux dernières avancées en apprentissage profond, ce qui a conduit à l'uniformisation de l'extraction des données au travers des communautés. Cependant, les architectures neuronales multimodales n'en sont qu'à leurs premiers balbutiements et l’apprentissage par renforcement profond est encore souvent restreint à des environnements limités. Idéalement, nous aimerions pourtant développer des modèles multimodaux et interactifs afin qu’ils puissent correctement appréhender la complexité du monde réel. Dans cet objectif, cette thèse s’attache à la compréhension du langage combiné à la vision pour trois raisons : (i) ce sont deux modalités longuement étudiées aux travers des différentes communautés scientifiques (ii) nous pouvons bénéficier des dernières avancées en apprentissage profond pour les modèles de langues et de vision (iii) l’interaction entre l’apprentissage du langage et notre perception a été validé en science cognitives. Ainsi, nous avons conçu le jeu GuessWhat?! (KéZaKo) afin d’évaluer la compréhension de langue combiné à la vision de nos modèles : deux joueurs doivent ainsi localiser un objet caché dans une image en posant une série de questions. Nous introduisons ensuite le principe de modulation comme un nouveau module d’apprentissage profond multimodal. Nous montrons qu’une telle approche permet de fusionner efficacement des représentations visuelles et langagières en prenant en compte la structure hiérarchique propre aux réseaux de neurones. Enfin, nous explorons comment l'apprentissage par renforcement permet l’apprentissage de la langue et cimente l'apprentissage des représentations multimodales sous-jacentes. Nous montrons qu’un tel apprentissage interactif conduit à des stratégies langagières valides mais donne lieu à de nouvelles problématiques de recherche
While our representation of the world is shaped by our perceptions, our languages, and our interactions, they have traditionally been distinct fields of study in machine learning. Fortunately, this partitioning started opening up with the recent advents of deep learning methods, which standardized raw feature extraction across communities. However, multimodal neural architectures are still at their beginning, and deep reinforcement learning is often limited to constrained environments. Yet, we ideally aim to develop large-scale multimodal and interactive models towards correctly apprehending the complexity of the world. As a first milestone, this thesis focuses on visually grounded language learning for three reasons (i) they are both well-studied modalities across different scientific fields (ii) it builds upon deep learning breakthroughs in natural language processing and computer vision (ii) the interplay between language and vision has been acknowledged in cognitive science. More precisely, we first designed the GuessWhat?! game for assessing visually grounded language understanding of the models: two players collaborate to locate a hidden object in an image by asking a sequence of questions. We then introduce modulation as a novel deep multimodal mechanism, and we show that it successfully fuses visual and linguistic representations by taking advantage of the hierarchical structure of neural networks. Finally, we investigate how reinforcement learning can support visually grounded language learning and cement the underlying multimodal representation. We show that such interactive learning leads to consistent language strategies but gives raise to new research issues

In this work, we make a simple but important observation questions about images often contain premises -- objects and relationships implied by the question -- and that reasoning about premises can help Visual Question Answering (VQA) models respond more intelligently to irrelevant or previously unseen questions. When presented with a question that is irrelevant to an image, state-of-the-art VQA models will still answer based purely on learned language biases, resulting in nonsensical or even misleading answers. We note that a visual question is irrelevant to an image if at least one of its premises is false (i.e. not depicted in the image). We leverage this observation to construct a dataset for Question Relevance Prediction and Explanation (QRPE) by searching for false premises. We train novel irrelevant question detection models and show that models that reason about premises consistently outperform models that do not. We also find that forcing standard VQA models to reason about premises during training can lead to improvements on tasks requiring compositional reasoning.
Master of Science

A natural application of artificial intelligence is to help blind people overcome their daily visual challenges through AI-based assistive technologies. In this regard, one of the most promising tasks is Visual Question Answering (VQA): the model is presented with an image and a question about this image. It must then predict the correct answer. Recently has been introduced the VizWiz dataset, a collection of images and questions originating from blind people. Being the first VQA dataset deriving from a natural setting, VizWiz presents many limitations and peculiarities. More specifically, the characteristics observed are the high uncertainty of the answers, the conversational aspect of questions, the relatively small size of the datasets and ultimately, the imbalance between answerable and unanswerable classes. These characteristics could be observed, individually or jointly, in other VQA datasets, resulting in a burden when solving the VQA task. Particularly suitable to address these aspects of the data are data science pre-processing techniques. Therefore, to provide a solid contribution to the VQA task, we answered the research question “Can data science pre-processing techniques improve the VQA task?” by proposing and studying the effects of four different pre-processing techniques. To address the high uncertainty of answers we employed a pre-processing step in which it is computed the uncertainty of each answer and used this measure to weight the soft scores of our model during training. The adoption of an “uncertainty-aware” training procedure boosted the predictive accuracy of our model of 10% providing a new state-of-the-art when evaluated on the test split of the VizWiz dataset. In order to overcome the limited amount of data, we designed and tested a new pre-processing procedure able to augment the training set and almost double its data points by computing the cosine similarity between answers representation. We addressed also the conversational aspect of questions collected from real world verbal conversations by proposing an alternative question pre-processing pipeline in which conversational terms are removed. This led in a further improvement: from a predictive accuracy of 0.516 with the standard question processing pipeline, we were able to achieve 0.527 predictive accuracy when employing the new pre-processing pipeline. Ultimately, we addressed the imbalance between answerable and unanswerable classes when predicting the answerability of a visual question. We tested two standard pre-processing techniques to adjust the dataset class distribution: oversampling and undersampling. Oversampling provided an albeit small improvement in both average precision and F1 score.
En naturlig tillämpning av artificiell intelligens är att hjälpa blinda med deras dagliga visuella utmaningar genom AI-baserad hjälpmedelsteknik. I detta avseende, är en av de mest lovande uppgifterna Visual Question Answering (VQA): modellen presenteras med en bild och en fråga om denna bild, och måste sedan förutspå det korrekta svaret. Nyligen introducerades VizWiz-datamängd, en samling bilder och frågor till dessa från blinda personer. Då detta är det första VQA-datamängden som härstammar från en naturlig miljö, har det många begränsningar och särdrag. Mer specifikt är de observerade egenskaperna: hög osäkerhet i svaren, informell samtalston i frågorna, relativt liten datamängd och slutligen obalans mellan svarbara och icke svarbara klasser. Dessa egenskaper kan även observeras, enskilda eller tillsammans, i andra VQA-datamängd, vilket utgör särskilda utmaningar vid lösning av VQA-uppgiften. Särskilt lämplig för att hantera dessa aspekter av data är förbehandlingsteknik från området data science. För att bidra till VQA-uppgiften, svarade vi därför på frågan “Kan förbehandlingstekniker från området data science bidra till lösningen av VQA-uppgiften?” genom att föreslå och studera effekten av fyra olika förbehandlingstekniker. För att hantera den höga osäkerheten i svaren använde vi ett förbehandlingssteg där vi beräknade osäkerheten i varje svar och använde detta mått för att vikta modellens utdata-värden under träning. Användandet av en ”osäkerhetsmedveten” träningsprocedur förstärkte den förutsägbara noggrannheten hos vår modell med 10%. Med detta nådde vi ett toppresultat när modellen utvärderades på testdelen av VizWiz-datamängden. För att övervinna problemet med den begränsade mängden data, konstruerade och testade vi en ny förbehandlingsprocedur som nästan dubblerar datapunkterna genom att beräkna cosinuslikheten mellan svarens vektorer. Vi hanterade även problemet med den informella samtalstonen i frågorna, som samlats in från den verkliga världens verbala konversationer, genom att föreslå en alternativ väg att förbehandla frågorna, där samtalstermer är borttagna. Detta ledde till en ytterligare förbättring: från en förutsägbar noggrannhet på 0.516 med det vanliga sättet att bearbeta frågorna kunde vi uppnå 0.527 prediktiv noggrannhet vid användning av det nya sättet att förbehandla frågorna. Slutligen hanterade vi obalansen mellan svarbara och icke svarbara klasser genom att förutse om en visuell fråga har ett möjligt svar. Vi testade två standard-förbehandlingstekniker för att justeradatamängdens klassdistribution: översampling och undersampling. Översamplingen gav en om än liten förbättring i både genomsnittlig precision och F1-poäng.

La quantité d'images présentes sur internet augmente considérablement, et il est nécessaire de développer des techniques permettant le traitement automatique de ces contenus. Alors que les méthodes de reconnaissance visuelle sont de plus en plus évoluées, la communauté scientifique s'intéresse désormais à des systèmes aux capacités de raisonnement plus poussées. Dans cette thèse, nous nous intéressons au Visual Question Answering (VQA), qui consiste en la conception de systèmes capables de répondre à une question portant sur une image. Classiquement, ces architectures sont conçues comme des systèmes d'apprentissage automatique auxquels on fournit des images, des questions et leur réponse. Ce problème difficile est habituellement abordé par des techniques d'apprentissage profond. Dans la première partie de cette thèse, nous développons des stratégies de fusion multimodales permettant de modéliser des interactions entre les représentations d'image et de question. Nous explorons des techniques de fusion bilinéaire, et assurons l'expressivité et la simplicité des modèles en utilisant des techniques de factorisation tensorielle. Dans la seconde partie, on s'intéresse au raisonnement visuel qui encapsule ces fusions. Après avoir présenté les schémas classiques d'attention visuelle, nous proposons une architecture plus avancée qui considère les objets ainsi que leurs relations mutuelles. Tous les modèles sont expérimentalement évalués sur des jeux de données standards et obtiennent des résultats compétitifs avec ceux de la littérature
The quantity of images that populate the Internet is dramatically increasing. It becomes of critical importance to develop the technology for a precise and automatic understanding of visual contents. As image recognition systems are becoming more and more relevant, researchers in artificial intelligence now seek for the next generation vision systems that can perform high-level scene understanding. In this thesis, we are interested in Visual Question Answering (VQA), which consists in building models that answer any natural language question about any image. Because of its nature and complexity, VQA is often considered as a proxy for visual reasoning. Classically, VQA architectures are designed as trainable systems that are provided with images, questions about them and their answers. To tackle this problem, typical approaches involve modern Deep Learning (DL) techniques. In the first part, we focus on developping multi-modal fusion strategies to model the interactions between image and question representations. More specifically, we explore bilinear fusion models and exploit concepts from tensor analysis to provide tractable and expressive factorizations of parameters. These fusion mechanisms are studied under the widely used visual attention framework: the answer to the question is provided by focusing only on the relevant image regions. In the last part, we move away from the attention mechanism and build a more advanced scene understanding architecture where we consider objects and their spatial and semantic relations. All models are thoroughly experimentally evaluated on standard datasets and the results are competitive with the literature

Learning and reasoning with common sense is a challenging problem in Artificial Intelligence (AI). Humans have the remarkable ability to interpret images and text from different perspectives in multiple modalities, and to use large amounts of commonsense knowledge while performing visual or textual tasks. Inspired by that ability, we approach commonsense learning as leveraging perspectives from multiple modalities for images and text in the context of vision and language tasks. Given a target task (e.g., textual reasoning, matching images with captions), our system first represents input images and text in multiple modalities (e.g., vision, text, abstract scenes and facts). Those modalities provide different perspectives to interpret the input images and text. And then based on those perspectives, the system performs reasoning to make a joint prediction for the target task. Surprisingly, we show that interpreting textual assertions and scene descriptions in the modality of abstract scenes improves performance on various textual reasoning tasks, and interpreting images in the modality of Visual Question Answering improves performance on caption retrieval, which is a visual reasoning task. With grounding, imagination and question-answering approaches to interpret images and text in different modalities, we show that learning commonsense knowledge from multiple modalities effectively improves the performance of downstream vision and language tasks, improves interpretability of the model and is able to make more efficient use of training data. Complementary to the model aspect, we also study the data aspect of commonsense learning in vision and language. We study active learning for Visual Question Answering (VQA) where a model iteratively grows its knowledge through querying informative questions about images for answers. Drawing analogies from human learning, we explore cramming (entropy), curiosity-driven (expected model change), and goal-driven (expected error reduction) active learning approaches, and propose a new goal-driven scoring function for deep VQA models under the Bayesian Neural Network framework. Once trained with a large initial training set, a deep VQA model is able to efficiently query informative question-image pairs for answers to improve itself through active learning, saving human effort on commonsense annotations.
Ph. D.

Visual Question Answering (VQA) models should have both high robustness and accuracy. Unfortunately, most of the current VQA research only focuses on accuracy because there is a lack of proper methods to measure the robustness of VQA models. There are two main modules in our algorithm. Given a natural language question about an image, the first module takes the question as input and then outputs the ranked basic questions, with similarity scores, of the main given question. The second module takes the main question, image and these basic questions as input and then outputs the text-based answer of the main question about the given image. We claim that a robust VQA model is one, whose performance is not changed much when related basic questions as also made available to it as input. We formulate the basic questions generation problem as a LASSO optimization, and also propose a large scale Basic Question Dataset (BQD) and Rscore (novel robustness measure), for analyzing the robustness of VQA models. We hope our BQD will be used as a benchmark for to evaluate the robustness of VQA models, so as to help the community build more robust and accurate VQA models.

Each time we ask for an object, describe a scene, follow directions or read a document containing images or figures, we are converting information between visual and linguistic representations. Indeed, for many tasks it is essential to reason jointly over visual and linguistic information. People do this with ease, typically without even noticing. Intelligent systems that perform useful tasks in unstructured situations, and interact with people, will also require this ability. In this thesis, we focus on the joint modelling of visual and linguistic information using deep neural networks. We begin by considering the challenging problem of automatically describing the content of an image in natural language, i.e., image captioning. Although there is considerable interest in this task, progress is hindered by the difficulty of evaluating the generated captions. Our first contribution is a new automatic image caption evaluation metric that measures the quality of generated captions by analysing their semantic content. Extensive evaluations across a range of models and datasets indicate that our metric, dubbed SPICE, shows high correlation with human judgements. Armed with a more effective evaluation metric, we address the challenge of image captioning. Visual attention mechanisms have been widely adopted in image captioning and visual question answering (VQA) architectures to facilitate fine-grained visual processing. We extend existing approaches by proposing a bottom-up and top-down attention mechanism that enables attention to be focused at the level of objects and other salient image regions, which is the natural basis for attention to be considered. Applying this approach to image captioning we achieve state of the art results on the COCO test server. Demonstrating the broad applicability of the method, applying the same approach to VQA we obtain first place in the 2017 VQA Challenge. Despite these advances, recurrent neural network (RNN) image captioning models typically do not generalise well to out-of-domain images containing novel scenes or objects. This limitation severely hinders the use of these models in real applications. To address this problem, we propose constrained beam search, an approximate search algorithm that enforces constraints over RNN output sequences. Using this approach, we show that existing RNN captioning architectures can take advantage of side information such as object detector outputs and ground-truth image annotations at test time, without retraining. Our results significantly outperform previous approaches that incorporate the same information into the learning algorithm, achieving state of the art results for out-of-domain captioning on COCO. Last, to enable and encourage the application of vision and language methods to problems involving embodied agents, we present the Matterport3D Simulator, a large-scale interactive reinforcement learning environment constructed from densely-sampled panoramic RGB-D images of 90 real buildings. Using this simulator, which can in future support a range of embodied vision and language tasks, we collect the first benchmark dataset for visually-grounded natural language navigation in real buildings. We investigate the difficulty of this task, and particularly the difficulty of operating in unseen environments, using several baselines and a sequence-to-sequence model based on methods successfully applied to other vision and language tasks.

abstract: Compressive sensing theory allows to sense and reconstruct signals/images with lower sampling rate than Nyquist rate. Applications in resource constrained environment stand to benefit from this theory, opening up many possibilities for new applications at the same time. The traditional inference pipeline for computer vision sequence reconstructing the image from compressive measurements. However,the reconstruction process is a computationally expensive step that also provides poor results at high compression rate. There have been several successful attempts to perform inference tasks directly on compressive measurements such as activity recognition. In this thesis, I am interested to tackle a more challenging vision problem - Visual question answering (VQA) without reconstructing the compressive images. I investigate the feasibility of this problem with a series of experiments, and I evaluate proposed methods on a VQA dataset and discuss promising results and direction for future work.
Dissertation/Thesis
Masters Thesis Computer Engineering 2017

abstract: Visual Question Answering (VQA) is a new research area involving technologies ranging from computer vision, natural language processing, to other sub-fields of artificial intelligence such as knowledge representation. The fundamental task is to take as input one image and one question (in text) related to the given image, and to generate a textual answer to the input question. There are two key research problems in VQA: image understanding and the question answering. My research mainly focuses on developing solutions to support solving these two problems. In image understanding, one important research area is semantic segmentation, which takes images as input and output the label of each pixel. As much manual work is needed to label a useful training set, typical training sets for such supervised approaches are always small. There are also approaches with relaxed labeling requirement, called weakly supervised semantic segmentation, where only image-level labels are needed. With the development of social media, there are more and more user-uploaded images available on-line. Such user-generated content often comes with labels like tags and may be coarsely labelled by various tools. To use these information for computer vision tasks, I propose a new graphic model by considering the neighborhood information and their interactions to obtain the pixel-level labels of the images with only incomplete image-level labels. The method was evaluated on both synthetic and real images. In question answering, my research centers on best answer prediction, which addressed two main research topics: feature design and model construction. In the feature design part, most existing work discussed how to design effective features for answer quality / best answer prediction. However, little work mentioned how to design features by considering the relationship between answers of one given question. To fill this research gap, I designed new features to help improve the prediction performance. In the modeling part, to employ the structure of the feature space, I proposed an innovative learning-to-rank model by considering the hierarchical lasso. Experiments with comparison with the state-of-the-art in the best answer prediction literature have confirmed that the proposed methods are effective and suitable for solving the research task.
Dissertation/Thesis
Doctoral Dissertation Computer Science 2017

碩士
國立交通大學
多媒體工程研究所
107
This thesis conducts a study of learning algorithms to address noisy label issues inherent in Visual Question Answering (VQA) tasks. The noisy labelling in VQA tasks refers to the phenomenon of possibly collecting different answers to an image-question pair from different human subjects. This often arises because some image-question pairs may create an ambiguous context that leads to indefinite answers. When trained with such noisy supervision, the performance of the VQA model suffers. To address noisy label issues, we first survey three mainstream algorithms for learning from noisy labels, including (1) loss-correction, (2) label cleansing and (3) graphical models. We then implement these algorithms based on a dual attention VQA network (which we call the base VQA model) and test their performance on VirginiaTech VQA dataset. Experimental results show that (1) the performances of the loss-correction algorithms rely heavily on accurate estimation of label transition probabilities due to noise or accurate detection of noise level, that (2) the label cleansing algorithms require enough verified labels to perform effectively, and that (3) the graphical models need to differentiate the noise level of each QA input to work well. In addition, the capability of the base VQA model can have a profound effect on the performances of these noisy label learning algorithms.

Humans have amazing visual perception which allows them to comprehend what the eyes see. In the core of human visual perception, lies the ability to translate visual information and link the visual information with linguistic cues from natural language. Visual reasoning and image understanding is a result of superior visual perception where one is able to comprehend visual and linguistic information and navigate these two domains seamlessly. The premise of Visual Question Answering (VQA) is to challenge an Artificial Intelligent (AI) agent by asking it to predict an answer for a natural language question about an image. By doing so, it evaluates its ability in the three major components of visual reasoning, first, simultaneous extraction of visual features from the image and semantic features from the question, second, joint processing of the multimodal features (visual and semantic), and third, learning to recognize regions in the image that are important to answer the question. In this thesis, we investigate how an AI agent can achieve human like visual reasoning and image understanding ability with superior visual perception, and is able to link linguistic cues with visual information when tasked with Visual Question Answering (VQA). Based on the observation that humans tend to ask questions about everyday objects and its attributes in context of the image, we developed a Reciprocal Attention Fusion (RAF) model, first of its kind, where the AI agent learns to simultaneously identify salient image regions of arbitrary shape and size, and rectangular object bounding boxes, for answering the question. We demonstrated that by combining these multilevel visual features and learning to identify image- and object-level attention map, our model learns to identify important visual cues for answering the question; thus achieving state-of-the art performance on several large scale VQA dataset. Further, we hypothesized that for achieving even better reasoning, a VQA model needs to attend all objects along with the objects deemed important by the question-driven attention mechanism. We developed a Question Agnostic Attention (QAA) model that forces any VQA model to consider all objects in the image along with their learned attention representing, which in turn results in better generalisation across and different high level reasoning task (i.e. counting, relative position), supporting our hypothesis. Furthermore, humans learn to identify relationships between object and describe them with semantic labels (e.g. in front of, seating) to get a holistic understanding of the image. We developed a semantic parser that generate linguistic features from subject-relationship-predicate triplets, and proposed an VQA model to incorporate this relationship parser on top of existing reasoning mechanism. This way we are able guide the VQA model to convert visual relationships to linguistic features, much like humans, and use it generate a answer which requires much higher reasoning than only identifying objects. In summary, in this thesis, we endeavour to improve the visual perception of Visual Linguistic AI agents by imitating human reasoning and image understanding process. It investigates how AI agents can incorporate different level of visual attention, learn to use high level linguistic cues as relationship labels, make use of transfer learning to reason about the unknown and also prove design recommendation to building such system. We hope our effort can help the community build better Visual Linguistic AI agents the can comprehend what the camera sees.

Visual Question Answering (VQA) is a recently proposed multimodal task in the general area of machine learning. The input to this task consists of a single image and an associated natural language question, and the output is the answer to that question. In this thesis we propose two incremental modifications to an existing model which won the VQA Challenge in 2016 using multimodal compact bilinear pooling (MCB), a novel way of combining modalities. First, we added the language attention mechanism, and on top of that we introduce an image attention mechanism focusing on objects detected in the image ("region attention"). We also experiment with ways of combining these in a single end- to-end model. The thesis describes the MCB model and our extensions and their two different implementations, and evaluates them on the original VQA challenge dataset for direct comparison with the original work. 1

Enabling machines to solve computer vision tasks with natural language components can greatly improve human interaction with computers. In this thesis, we address vision and language tasks with deep learning methods that explicitly localize relevant visual evidence. Spatial evidence localization in images enhances the interpretability of the model, while temporal localization in video is necessary to remove irrelevant content. We apply our methods to various vision and language tasks, including visual question answering, temporal activity detection, dense video captioning and cross-modal retrieval. First, we tackle the problem of image question answering, which requires the model to predict answers to questions posed about images. We design a memory network with a question-guided spatial attention mechanism which assigns higher weights to regions that are more relevant to the question. The visual evidence used to derive the answer can be shown by visualizing the attention weights in images. We then address the problem of localizing temporal evidence in videos. For most language/vision tasks, only part of the video is relevant to the linguistic component, so we need to detect these relevant events in videos. We propose an end-to-end model for temporal activity detection, which can detect arbitrary length activities by coordinate regression with respect to anchors and contains a proposal stage to filter out background segments, saving computation time. We further extend activity category detection to event captioning, which can express richer semantic meaning compared to a class label. This derives the problem of dense video captioning, which involves two sub-problems: localizing distinct events in long video and generating captions for the localized events. We propose an end-to-end hierarchical captioning model with vision and language context modeling in which the captioning training affects the activity localization. Lastly, the task of text-to-clip video retrieval requires one to localize the specified query instead of detecting and captioning all events. We propose a model based on the early fusion of words and visual features, outperforming standard approaches which embed the whole sentence before performing late feature fusion. Furthermore, we use queries to regulate the proposal network to generate query related proposals. In conclusion, our proposed visual localization mechanism applies across a variety of vision and language tasks and achieves state-of-the-art results. Together with the inference module, our work can contribute to solving other tasks such as video question answering in future research.

En utilisant la méthodologie de l'apprentissage profond qui préconise de s'appuyer davantage sur des données et des modèles neuronaux flexibles plutôt que sur les connaissances de l'expert dans le domaine, la communauté de recherche a récemment réalisé des progrès remarquables dans la compréhension et la génération du langue naturel. Néanmoins, il reste difficile de savoir si une simple extension des méthodes d'apprentissage profond existantes sera suffisante pour atteindre l'objectif d'utiliser le langage naturel pour l'interaction homme-machine. Nous nous concentrons sur deux aspects connexes dans lesquels les méthodes actuelles semblent nécessiter des améliorations majeures. Le premier de ces aspects est l'inefficacité statistique des systèmes d'apprentissage profond: ils sont connus pour nécessiter de grandes quantités de données pour bien fonctionner. Le deuxième aspect est leur capacité limitée à généraliser systématiquement, à savoir à comprendre le langage dans des situations où la distribution des données change mais les principes de syntaxe et de sémantique restent les mêmes. Dans cette thèse, nous présentons quatre études de cas dans lesquelles nous cherchons à apporter plus de clarté concernant l'efficacité statistique susmentionnée et les aspects de généralisation systématique des approches d'apprentissage profond de la compréhension des langues, ainsi qu'à faciliter la poursuite des travaux sur ces sujets. Afin de séparer le problème de la représentation des connaissances du monde réel du problème de l'apprentissage d'une langue, nous menons toutes ces études en utilisant des langages synthétiques ancrés dans des environnements visuels simples. Dans le premier article, nous étudions comment former les agents à suivre des instructions compositionnelles dans des environnements avec une forme de supervision restreinte. À savoir pour chaque instruction et configuration initiale de l'environnement, nous ne fournissons qu'un état cible au lieu d'une trajectoire complète avec des actions à toutes les étapes. Nous adaptons les méthodes d'apprentissage adversariel par imitation à ce paramètre et démontrons qu'une telle forme restreinte de données est suffisante pour apprendre les significations compositionelles des instructions. Notre deuxième article se concentre également sur des agents qui apprennent à exécuter des instructions. Nous développons la plateforme BabyAI pour faciliter des études plus approfondies et plus rigoureuses de ce cadre d'apprentissage. La plateforme fournit une langue BabyAI compositionnelle avec $10 ^ {19}$ instructions, dont la sémantique est précisément définie dans un environnement partiellement observable. Nous rapportons des résultats de référence sur la quantité de supervision nécessaire pour enseigner à l'agent certains sous-ensembles de la langue BabyAI avec différentes méthodes de formation, telles que l'apprentissage par renforcement et l'apprentissage par imitation. Dans le troisième article, nous étudions la généralisation systématique des modèles de réponse visuelle aux questions (VQA). Dans le scénario VQA, le système doit répondre aux questions compositionelles sur les images. Nous construisons un ensemble de données de questions spatiales sur les paires d'objets et évaluons la performance des différents modèles sur les questions concernant les paires d'objets qui ne se sont jamais produites dans la même question dans la distribution d'entraînement. Nous montrons que les modèles dans lesquels les significations des mots sont représentés par des modules séparés qui effectuent des calculs indépendants généralisent beaucoup mieux que les modèles dont la conception n'est pas explicitement modulaire. Cependant, les modèles modulaires ne généralisent bien que lorsque les modules sont connectés dans une disposition appropriée, et nos expériences mettent en évidence les défis de l'apprentissage de la disposition par un apprentissage de bout en bout sur la distribution d'entraînement. Dans notre quatrième et dernier article, nous étudions également la généralisation des modèles VQA à des questions en dehors de la distribution d'entraînement, mais cette fois en utilisant le jeu de données CLEVR, utilisé pour les questions complexes sur des scènes rendues en 3D. Nous générons de nouvelles questions de type CLEVR en utilisant des références basées sur la similitude (par exemple `` la balle qui a la même couleur que ... '') dans des contextes qui se produisent dans les questions CLEVR mais uniquement avec des références basées sur la localisation (par exemple `` le balle qui est à gauche de ... ''). Nous analysons la généralisation avec zéro ou quelques exemples de CLOSURE après un entraînement sur CLEVR pour un certain nombre de modèles existants ainsi qu'un nouveau modèle.
By using the methodology of deep learning that advocates relying more on data and flexible neural models rather than on the expert's knowledge of the domain, the research community has recently achieved remarkable progress in natural language understanding and generation. Nevertheless, it remains unclear whether simply scaling up existing deep learning methods will be sufficient to achieve the goal of using natural language for human-computer interaction. We focus on two related aspects in which current methods appear to require major improvements. The first such aspect is the data inefficiency of deep learning systems: they are known to require extreme amounts of data to perform well. The second aspect is their limited ability to generalize systematically, namely to understand language in situations when the data distribution changes yet the principles of syntax and semantics remain the same. In this thesis, we present four case studies in which we seek to provide more clarity regarding the aforementioned data efficiency and systematic generalization aspects of deep learning approaches to language understanding, as well as to facilitate further work on these topics. In order to separate the problem of representing open-ended real-world knowledge from the problem of core language learning, we conduct all these studies using synthetic languages that are grounded in simple visual environments. In the first article, we study how to train agents to follow compositional instructions in environments with a restricted form of supervision. Namely for every instruction and initial environment configuration we only provide a goal-state instead of a complete trajectory with actions at all steps. We adapt adversarial imitation learning methods to this setting and demonstrate that such a restricted form of data is sufficient to learn compositional meanings of the instructions. Our second article also focuses on instruction following. We develop the BabyAI platform to facilitate further, more extensive and rigorous studies of this setup. The platform features a compositional Baby language with $10^{19}$ instructions, whose semantics is precisely defined in a partially-observable gridworld environment. We report baseline results on how much supervision is required to teach the agent certain subsets of Baby language with different training methods, such as reinforcement learning and imitation learning. In the third article we study systematic generalization of visual question answering (VQA) models. In the VQA setting the system must answer compositional questions about images. We construct a dataset of spatial questions about object pairs and evaluate how well different models perform on questions about pairs of objects that never occured in the same question in the training distribution. We show that models in which word meanings are represented by separate modules that perform independent computation generalize much better than models whose design is not explicitly modular. The modular models, however, generalize well only when the modules are connected in an appropriate layout, and our experiments highlight the challenges of learning the layout by end-to-end learning on the training distribution. In our fourth and final article we also study generalization of VQA models to questions outside of the training distribution, but this time using the popular CLEVR dataset of complex questions about 3D-rendered scenes as the platform. We generate novel CLEVR-like questions by using similarity-based references (e.g. ``the ball that has the same color as ...'') in contexts that occur in CLEVR questions but only with location-based references (e.g. ``the ball that is to the left of ...''). We analyze zero- and few- shot generalization to CLOSURE after training on CLEVR for a number of existing models as well as a novel one.