The mission of the research presented in this thesis is to give computers the power to sense and react to human activities. Without the ability to sense the surroundings and understand what humans are doing, computers will not be able to provide active, timely, appropriate, and considerate services to the humans. To accomplish this mission, the work stands on the shoulders of two giants: Machine learning and ubiquitous computing. Because of the ubiquity of sensor-enabled mobile and wearable devices, there has been an emerging opportunity to sense, learn, and infer human activities from the sensor data by leveraging state-of-the-art machine learning algorithms. While having shown promising results in human activity recognition, most existing approaches using supervised or semi-supervised learning have two fundamental problems. Firstly, most existing approaches require a large set of labeled sensor data for every target class, which requires a costly effort from human annotators. Secondly, an unseen new activity cannot be recognized if no training samples of that activity are available in the dataset. In light of these problems, a new approach in this area is proposed in our research. This thesis presents our novel approach to address the problem of human activity recognition when few or no training samples of the target activities are available. The main hypothesis is that the problem can be solved by the proposed NuActiv activity recognition framework, which consists of modeling the hierarchical and sequential structure of human activities, as well as bringing humans in the loop of model training. By injecting human knowledge about the hierarchical nature of human activities, a semantic attribute representation and a two-layer attribute-based learning approach are designed. To model the sequential structure, a probabilistic graphical model is further proposed to take into account the temporal dependency of activities and attributes. Finally, an active learning algorithm is developed to reinforce the recognition accuracy using minimal user feedback. The hypothesis and approaches presented in this thesis are validated by two case studies and real-world experiments on exercise activities and daily life activities. Experimental results show that the NuActiv framework can effectively recognize unseen new activities even without any training data, with up to 70-80% precision and recall rate. It also outperforms supervised learning with limited labeled data for the new classes. The results significantly advance the state of the art in human activity recognition, and represent a promising step towards bridging the gap between computers and humans.

Nous assistons actuellement à une explosion de la quantité des données visuelles. Par exemple, plusieurs millions de photos sont partagées quotidiennement sur les réseaux sociaux. Les méthodes d'interprétation d'images vise à faciliter l'accès à ces données visuelles, d'une manière sémantiquement compréhensible. Dans ce manuscrit, nous définissons certains buts détaillés qui sont intéressants pour les taches d'interprétation d'images, telles que la classification ou la recherche d'images, que nous considérons dans les trois chapitres principaux. Tout d'abord, nous visons l'exploitation de la nature multimodale de nombreuses bases de données, pour lesquelles les documents sont composés d'images et de descriptions textuelles. Dans ce but, nous définissons des similarités entre le contenu visuel d'un document, et la description textuelle d'un autre document. Ces similarités sont calculées en deux étapes, tout d'abord nous trouvons les voisins visuellement similaires dans la base multimodale, puis nous utilisons les descriptions textuelles de ces voisins afin de définir une similarité avec la description textuelle de n'importe quel document. Ensuite, nous présentons une série de modèles structurés pour la classification d'images, qui encodent explicitement les interactions binaires entre les étiquettes (ou labels). Ces modèles sont plus expressifs que des prédicateurs d'étiquette indépendants, et aboutissent à des prédictions plus fiables, en particulier dans un scenario de prédiction interactive, où les utilisateurs fournissent les valeurs de certaines des étiquettes d'images. Un scenario interactif comme celui-ci offre un compromis intéressant entre la précision, et l'effort d'annotation manuelle requis. Nous explorons les modèles structurés pour la classification multi-étiquette d'images, pour la classification d'image basée sur les attributs, et pour l'optimisation de certaines mesures de rang spécifiques. Enfin, nous explorons les classifieurs par k plus proches voisins, et les classifieurs par plus proche moyenne, pour la classification d'images à grande échelle. Nous proposons des méthodes d'apprentissage de métrique efficaces pour améliorer les performances de classification, et appliquons ces méthodes à une base de plus d'un million d'images d'apprentissage, et d'un millier de classes. Comme les deux méthodes de classification permettent d'incorporer des classes non vues pendant l'apprentissage à un coût presque nul, nous avons également étudié leur performance pour la généralisation. Nous montrons que la classification par plus proche moyenne généralise à partir d'un millier de classes, sur dix mille classes à un coût négligeable, et les performances obtenus sont comparables à l'état de l'art. / We are currently experiencing an exceptional growth of visual data, for example, millions of photos are shared daily on social-networks. Image understanding methods aim to facilitate access to this visual data in a semantically meaningful manner. In this dissertation, we define several detailed goals which are of interest for the image understanding tasks of image classification and retrieval, which we address in three main chapters. First, we aim to exploit the multi-modal nature of many databases, wherein documents consists of images with a form of textual description. In order to do so we define similarities between the visual content of one document and the textual description of another document. These similarities are computed in two steps, first we find the visually similar neighbors in the multi-modal database, and then use the textual descriptions of these neighbors to define a similarity to the textual description of any document. Second, we introduce a series of structured image classification models, which explicitly encode pairwise label interactions. These models are more expressive than independent label predictors, and lead to more accurate predictions. Especially in an interactive prediction scenario where a user provides the value of some of the image labels. Such an interactive scenario offers an interesting trade-off between accuracy and manual labeling effort. We explore structured models for multi-label image classification, for attribute-based image classification, and for optimizing for specific ranking measures. Finally, we explore k-nearest neighbors and nearest-class mean classifiers for large-scale image classification. We propose efficient metric learning methods to improve classification performance, and use these methods to learn on a data set of more than one million training images from one thousand classes. Since both classification methods allow for the incorporation of classes not seen during training at near-zero cost, we study their generalization performances. We show that the nearest-class mean classification method can generalize from one thousand to ten thousand classes at negligible cost, and still perform competitively with the state-of-the-art.

La construction d'algorithmes classifiant des images à grande échelle est devenue une t^ache essentielle du fait de la difficulté d'effectuer des recherches dans les immenses collections de données visuelles non-etiquetées présentes sur Internet. L'objetif est de classifier des images en fonction de leur contenu pour simplifier la gestion de telles bases de données. La classification d'images à grande échelle est un problème complexe, de par l'importance de la taille des ensembles de données, tant en nombre d'images qu'en nombre de classes. Certaines de ces classes sont dites "fine-grained" (sémantiquement proches les unes des autres) et peuvent même ne contenir aucun représentant étiqueté. Dans cette thèse, nous utilisons des représentations à l'état de l'art d'images et nous concentrons sur des méthodes d'apprentissage efficaces. Nos contributions sont (1) un banc d'essai d'algorithmes d'apprentissage pour la classification à grande échelle et (2) un nouvel algorithme basé sur l'incorporation d'étiquettes pour apprendre sur des données peu abondantes. En premier lieu, nous introduisons un banc d'essai d'algorithmes d'apprentissage pour la classification à grande échelle, dans un cadre entièrement supervisé. Il compare plusieurs fonctions objectifs pour apprendre des classifieurs linéaires, tels que "un contre tous", "multiclasse", "classement", "classement avec pondération" par descente de gradient stochastique. Ce banc d'essai se conclut en un ensemble de recommandations pour la classification à grande échelle. Avec une simple repondération des données, la stratégie "un contre tous" donne des performances meilleures que toutes les autres. Par ailleurs, en apprentissage en ligne, un pas d'apprentissage assez petit s'avère suffisant pour obtenir des résultats au niveau de l'état de l'art. Enfin, l'arrêt prématuré de la descente de gradient stochastique introduit une régularisation qui améliore la vitesse d'entraînement ainsi que la capacité de régularisation. Deuxièmement, face à des milliers de classes, il est parfois difficile de rassembler suffisamment de données d'entraînement pour chacune des classes. En particulier, certaines classes peuvent être entièrement dénuées d'exemples. En conséquence, nous proposons un nouvel algorithme adapté à ce scénario d'apprentissage dit "zero-shot". Notre algorithme utilise des données parallèles, comme les attributs, pour incorporer les classes dans un espace euclidien. Nous introduisons par ailleurs une fonction pour mesurer la compatibilité entre image et étiquette. Les paramètres de cette fonction sont appris en utilisant un objectif de type "ranking". Notre algorithme dépasse l'état de l'art pour l'apprentissage "zero-shot", et fait preuve d'une grande flexibilité en permettant d'incorporer d'autres sources d'information parallèle, comme des hiérarchies. Il permet en outre une transition sans heurt du cas "zero-shot" au cas où peu d'exemples sont disponibles. / Building algorithms that classify images on a large scale is an essential task due to the difficulty in searching massive amount of unlabeled visual data available on the Internet. We aim at classifying images based on their content to simplify the manageability of such large-scale collections. Large-scale image classification is a difficult problem as datasets are large with respect to both the number of images and the number of classes. Some of these classes are fine grained and they may not contain any labeled representatives. In this thesis, we use state-of-the-art image representations and focus on efficient learning methods. Our contributions are (1) a benchmark of learning algorithms for large scale image classification, and (2) a novel learning algorithm based on label embedding for learning with scarce training data. Firstly, we propose a benchmark of learning algorithms for large scale image classification in the fully supervised setting. It compares several objective functions for learning linear classifiers such as one-vs-rest, multiclass, ranking and weighted average ranking using the stochastic gradient descent optimization. The output of this benchmark is a set of recommendations for large-scale learning. We experimentally show that, online learning is well suited for large-scale image classification. With simple data rebalancing, One-vs-Rest performs better than all other methods. Moreover, in online learning, using a small enough step size with respect to the learning rate is sufficient for state-of-the-art performance. Finally, regularization through early stopping results in fast training and a good generalization performance. Secondly, when dealing with thousands of classes, it is difficult to collect sufficient labeled training data for each class. For some classes we might not even have a single training example. We propose a novel algorithm for this zero-shot learning scenario. Our algorithm uses side information, such as attributes to embed classes in a Euclidean space. We also introduce a function to measure the compatibility between an image and a label. The parameters of this function are learned using a ranking objective. Our algorithm outperforms the state-of-the-art for zero-shot learning. It is flexible and can accommodate other sources of side information such as hierarchies. It also allows for a smooth transition from zero-shot to few-shots learning.

[Background]: The advent of bidirectional encoder representation from trans- formers (BERT) language models (Devlin et al., 2018) and MS Marco, a large scale human-annotated dataset for machine reading comprehension (Bajaj et al., 2016) that made publicly available, led the field of information retrieval (IR) to experience a revolution (Lin et al., 2020). The retrieval model based on BERT of Nogueira and Cho (2019), by the time they published their paper, became the top entry in the MS Marco passage-reranking leaderboard, surpassing the previous state of the art by 27% in MRR@10. However, training such neural IR models for different domains than MS Marco is still hard because neural approaches often require a vast amount of training data to perform effectively, which is not always available. To address the problem of the shortage of labelled data a new line of research emerged, training neural models with weak supervision. In weak supervision, given an unlabelled dataset labels are generated automatically using an existing model and then a machine learning model is trained upon the artificial “weak“ data. In case of weak supervision for IR, the training dataset comes in the form of a tuple (query, passage). Dehghani et al. (2017) in their work used the AOL query logs (Pass et al., 2006), which is a set of millions of real web queries, and BM25 to retrieve the relevant passages for each of the user queries. A drawback with this approach is that it is hard to obtain query logs for every single different domain. [Objective]: This thesis proposes an intuitive approach for addressing the shortage of data in domains with limited or no data at all through transfer learning in the context of IR. We leverage Wikipedia’s structure for creating a Wikipedia-based generic IR training dataset for zero-shot neural models. [Method]: We create the “pseudo-queries“ by concatenating the titles of Wikipedia’s articles along with each of their title sections and we consider the associated section’s passage as the relevant passage of the pseudo-queries. All of our experiments are evaluated on a standard collection: MS Marco, which is a large scale web collection. For our zero-shot experiments, our proposed model, called “Wiki“, is a BERT model trained on the artificial Wikipedia-based dataset and the baseline is a default BERT model without any additional training. In our second line of experiments, we explore the benefits gained by pre-fine- tuning on the Wikipedia-based IR dataset and further fine-tuning on in-domain data. Our proposed model, "Wiki+Ma", is a BERT model pre-fine-tuned in the Wikipedia-based dataset and further fine-tuned in MS Marco, while the baseline is a BERT model fine-tuned only in MS Marco. [Results]: Results regarding our first experiments show that our BERT model trained on the Wikipedia-based IR dataset, called "Wiki", achieves a performance of 0.197 in MRR@10, which is about +10 points more in comparison to a BERT model with default weights; in addition, results in the development set indicate that the “Wiki“ model performs better than BERT model trained on in-domain data when the data is between 10k-50k instances. Results regarding our second line of experiments show that pre-fine-tuning on the Wikipedia-based IR dataset benefits later fine-tuning steps on in-domain data in terms of stability. [Conclusion]: Our findings suggest that transfer learning for IR tasks by leveraging the generic knowledge incorporated in Wikipedia is possible, though more experimentation is needed to understand its limitations in comparison with the traditional approaches such as the BM25.

L'utilisation d'architectures neuronales profondes associée à des innovations spécifiques telles que les méthodes adversarielles, l’entraînement préalable sur de grands ensembles de données et l'estimation de l'information mutuelle a permis, ces dernières années, de progresser rapidement dans de nombreuses tâches de vision par ordinateur complexes telles que la classification d'images de catégories préalablement inconnues (apprentissage zéro-coups), la génération de scènes ou la classification multimodale. Malgré ces progrès, il n’est pas certain que les méthodes actuelles d’apprentissage de représentations suffiront à atteindre une performance équivalente au niveau humain sur des tâches visuelles arbitraires et, de fait, cela pose des questions quant à la direction de la recherche future. Dans cette thèse, nous nous concentrerons sur deux aspects des représentations qui semblent nécessaires pour atteindre de bonnes performances en aval pour l'apprentissage des représentations : la localité et la compositionalité. La localité peut être comprise comme la capacité d'une représentation à retenir des informations locales. Ceci sera pertinent dans de nombreux cas, et bénéficiera particulièrement à la vision informatique, domaine dans lequel les images naturelles comportent intrinsèquement des informations locales, par exemple des parties pertinentes d’une image, des objets multiples présents dans une scène... D'autre part, une représentation compositionnelle peut être comprise comme une représentation qui résulte d'une combinaison de parties plus simples. Les réseaux neuronaux convolutionnels sont intrinsèquement compositionnels, et de nombreuses images complexes peuvent être considérées comme la composition de sous-composantes pertinentes : les objets et attributs individuels dans une scène, les attributs sémantiques dans l'apprentissage zéro-coups en sont deux exemples. Nous pensons que ces deux propriétés détiennent la clé pour concevoir de meilleures méthodes d'apprentissage de représentations. Dans cette thèse, nous présentons trois articles traitant de la localité et/ou de la compositionnalité, et de leur application à l'apprentissage de représentations pour des tâches visuelles complexes. Dans le premier article, nous introduisons des méthodes de mesure de la localité et de la compositionnalité pour les représentations d'images, et nous démontrons que les représentations locales et compositionnelles sont plus performantes dans l'apprentissage zéro-coups. Nous utilisons également ces deux notions comme base pour concevoir un nouvel algorithme d'apprentissage des représentations qui atteint des performances de pointe dans notre cadre expérimental, une variante de l'apprentissage "zéro-coups" plus difficile où les informations externes, par exemple un pré-entraînement sur d'autres ensembles de données d'images, ne sont pas autorisées. Dans le deuxième article, nous montrons qu'en encourageant un générateur à conserver des informations locales au niveau de l'objet, à l'aide d'un module dit de similarité de graphes de scène, nous pouvons améliorer les performances de génération de scènes. Ce modèle met également en évidence l'importance de la composition, car de nombreux composants fonctionnent individuellement sur chaque objet présent. Pour démontrer pleinement la portée de notre approche, nous effectuons une analyse détaillée et proposons un nouveau cadre pour évaluer les modèles de génération de scènes. Enfin, dans le troisième article, nous montrons qu'en encourageant une forte information mutuelle entre les représentations multimodales locales et globales des images médicales en 2D et 3D, nous pouvons améliorer la classification et la segmentation des images. Ce cadre général peut être appliqué à une grande variété de contextes et démontre les avantages non seulement de la localité, mais aussi de la compositionnalité, car les représentations multimodales sont combinées pour obtenir une représentation plus générale. / The use of deep neural architectures coupled with specific innovations such as adversarial methods, pre-training on large datasets and mutual information estimation has in recent years allowed rapid progress in many complex vision tasks such as zero-shot learning, scene generation, or multi-modal classification. Despite such progress, it is still not clear if current representation learning methods will be enough to attain human-level performance on arbitrary visual tasks, and if not, what direction should future research take. In this thesis, we will focus on two aspects of representations that seem necessary to achieve good downstream performance for representation learning: locality and compositionality. Locality can be understood as a representation's ability to retain local information. This will be relevant in many cases, and will specifically benefit computer vision where natural images inherently feature local information, i.e. relevant patches of an image, multiple objects present in a scene... On the other hand, a compositional representation can be understood as one that arises from a combination of simpler parts. Convolutional neural networks are inherently compositional, and many complex images can be seen as composition of relevant sub-components: individual objects and attributes in a scene, semantic attributes in zero-shot learning are two examples. We believe both properties hold the key to designing better representation learning methods. In this thesis, we present 3 articles dealing with locality and/or compositionality, and their application to representation learning for complex visual tasks. In the first article, we introduce ways of measuring locality and compositionality for image representations, and demonstrate that local and compositional representations perform better at zero-shot learning. We also use these two notions as the basis for designing class-matching deep info-max, a novel representation learning algorithm that achieves state-of-the-art performance on our proposed "Zero-shot from scratch" setting, a harder zero-shot setting where external information, e.g. pre-training on other image datasets is not allowed. In the second article, we show that by encouraging a generator to retain local object-level information, using a scene-graph similarity module, we can improve scene generation performance. This model also showcases the importance of compositionality as many components operate individually on each object present. To fully demonstrate the reach of our approach, we perform detailed analysis, and propose a new framework to evaluate scene generation models. Finally, in the third article, we show that encouraging high mutual information between local and global multi-modal representations of 2D and 3D medical images can lead to improvements in image classification and segmentation. This general framework can be applied to a wide variety of settings, and demonstrates the benefits of not only locality, but also of compositionality as multi-modal representations are combined to obtain a more general one.

The rapid growth of textual data in the digital age presents unique challenges in domain-specific text classification, particularly the scarcity of labeled data for many applications, due to expensive cost of manual labeling work. In this thesis, we explore the applicability of prompt-learning method, which is well-known for being suitable in few-shot scenarios and much less data-consuming, as an emerging alternative to traditional fine-tuning methods, for domain-specific text classification in the context of customer-agent interactions in the retail sector. Specifically, we implemented the entire prompt-learning pipeline for the classification task, and, our investigation encompasses various strategies of prompt-learning, including fixed-prompt language model tuning strategy and tuning-free prompting strategy, along with an examination of language model selection, few-shot sampling strategy, prompt template design, and verbalizer design. In this manner, we assessed the overall performance of the prompt-learning method in the classification task. Through a systematic evaluation, we demonstrate that with the fixed-prompt language model tuning strategy, based on relatively smaller language models (e.g. T5-base with around 220M parameters), prompt-learning can achieve competitive performance (close to 75% accuracy) even with limited labeled data (up to merely 15% of full data). And besides, with the tuning-free prompting strategy, based on a regular-size language model (e.g. FLAN-T5-large with around 770M parameters), the performance can be up to around 30% accuracy with detailed prompt templates and zero-shot setting (no extra training data involved). These results can offer valuable insights for researchers and practitioners working with domain-specific textual data, prompt-learning and few-shot / zero-shot learning. The findings of this thesis highlight the potential of prompt-learning as a practical solution for classification problems across diverse domains and set the stage for future research in this area.

Text-to-SQL, which involves translating natural language into Structured Query Language (SQL), is crucial for enabling broad access to structured databases without expert knowledge. However, designing models for such tasks is challenging due to numerous factors, including the presence of ’noise,’ such as ambiguous questions and syntactical errors. This thesis provides an in-depth analysis of the distribution and types of noise in the widely used BIRD-Bench benchmark and the impact of noise on models. While BIRD-Bench was created to model dirty and noisy database values, it was not created to contain noise and errors in the questions and gold queries. We found after a manual evaluation that noise in questions and gold queries are highly prevalent in the financial domain of the dataset, and a further analysis of the other domains indicate the presence of noise in other parts as well. The presence of incorrect gold SQL queries, which then generate incorrect gold answers, has a significant impact on the benchmark’s reliability. Surprisingly, when evaluating models on corrected SQL queries, zero-shot baselines surpassed the performance of state-of-the-art prompting methods. The thesis then introduces the concept of classifying noise in natural language questions, aiming to prevent the entry of noisy questions into text-to-SQL models and to annotate noise in existing datasets. Experiments using GPT-3.5 and GPT-4 on a manually annotated dataset demonstrated the viability of this approach, with classifiers achieving up to 0.81 recall and 80% accuracy. Additionally, the thesis explored the use of LLMs for automatically correcting faulty SQL queries. This showed a 100% success rate for specific query corrections, highlighting the potential for LLMs in improving dataset quality. We conclude that informative noise labels and reliable benchmarks are crucial to developing new Text-to-SQL methods that can handle varying types of noise.