Address parsing is the process of splitting an address string into its different address components, such as street name, street number, et cetera. Address parsing has been quite extensively researched and there exist some state-ofthe-art address parsing solutions, mostly unilingual. In more recent years research has emerged which focuses on multinational address parsing and deep architecture address parsers have been used to achieve state-of-the-art performance on multinational address data. However, training these deep architectures for address parsing requires a rather large amount of address data which is not always accessible. Generally within Natural Language Processing (NLP) data is difficult to come by and most of the NLP data available consists of data from about only 20 of the approximately 7000 languages spoken around the world, so-called high-resource languages. This also applies to address data, which can be difficult to come by for some of the so-called low-resource languages of the world for which little or no NLP data exists. To attempt to deal with the lack of address data availability for some of the less spoken languages of the world, the current project investigates the potential of FewShot Learning (FSL) for multinational address parsing. To investigate this, two few-shot transfer learning models are implemented, both implementations consist of a fine-tuned pre-trained language model (PTLM). The difference between the two models is the PTLM used, which were the multilingual language models mBERT and XLM-R, respectively. The two PTLMs are finetuned using a linear classifier layer to then be used as multinational address parsers. The two models are trained and their results are compared with a state-of-the-art multinational address parser, Deepparse, as well as with each other. Results show that the two models do not outperform Deepparse, but they do show promising results, not too far from what Deepparse achieves on holdout and zero-shot datasets. On a mix of low- and high-resource language address data, both models perform well and achieve over 96% on the overall F1-score. Out of the two models used for implementation, XLM-R achieves significantly better results than mBERT and can therefore be considered the more appropriate PTLM to use for multinational FSL address parsing. Based on these results the conclusion is that there is great potential for FSL within the field of multinational address parsing and that general FSL methods can be used and perform well on multinational address parsing tasks. / Adressavkodning är processen att dela upp en adresssträng i dess olika adresskomponenter såsom gatunamn, gatunummer, et cetera. Adressavkodning har undersökts ganska omfattande och det finns några toppmoderna adressavkodningslösningar, mestadels enspråkiga. Senaste åren har forskning fokuserad på multinationell adressavkodning börjat dyka upp och djupa arkitekturer för adressavkodning har använts för att uppnå toppmodern prestation på multinationell adressdata. Att träna dessa arkitekturer kräver dock en ganska stor mängd adressdata, vilket inte alltid är tillgängligt. Det är generellt svårt att få tag på data inom naturlig språkbehandling och majoriteten av den data som är tillgänglig består av data från endast 20 av de cirka 7000 språk som används runt om i världen, så kallade högresursspråk. Detta gäller även för adressdata, vilket kan vara svårt att få tag på för vissa av världens så kallade resurssnåla språk för vilka det finns lite eller ingen data för naturlig språkbehandling. För att försöka behandla denna brist på adressdata för några av världens mindre talade språk undersöker detta projekt om det finns någon potential för inlärning med få exempel för multinationell adressavkodning. För detta implementeras två modeller för överföringsinlärning med få exempel genom finjustering av förtränade språkmodeller. Skillnaden mellan de två modellerna är den förtränade språkmodellen som används, mBERT respektive XLM-R. Båda modellerna finjusteras med hjälp av ett linjärt klassificeringsskikt för att sedan användas som multinationella addressavkodare. De två modellerna tränas och deras resultat jämförs med en toppmodern multinationell adressavkodare, Deepparse. Resultaten visar att de två modellerna presterar båda sämre än Deepparse modellen, men de visar ändå lovande resultat, inte långt ifrån vad Deepparse uppnår för både holdout och zero-shot dataset. Vidare, så presterar båda modeller bra på en blandning av adressdata från låg- och högresursspråk och båda modeller uppnår över 96% övergripande F1-score. Av de två modellerna uppnår XLM-R betydligt bättre resultat än mBERT och kan därför anses vara en mer lämplig förtränad språkmodell att använda för multinationell inlärning med få exempel för addressavkodning. Utifrån dessa resultat dras slutsatsen att det finns stor potential för inlärning med få exempel inom området multinationall adressavkodning, samt att generella metoder för inlärning med få exempel kan användas och preseterar bra på multinationella adressavkodningsuppgifter.

To label words of interest into a predefined set of named entities have traditionally required a large amount of labeled in-domain data. Recently, the availability of pre-trained transformer-based language models have enabled multiple natural language processing problems to utilize transfer learning techniques to construct machine learning models with less task-specific labeled data. In this thesis, the impact of data augmentation when training a pre-trained transformer-based model to adapt to a named entity recognition task with few labeled sentences is explored. The experimental results indicate that data augmentation increases performance of the trained models, however the data augmentation is shown to have less impact when more labeled data is available. In conclusion, data augmentation has been shown to improve performance of pre-trained named entity recognition models when few labeled sentences are available for training. / Att kategorisera ord som tillhör någon av en mängd förangivna entiteter har traditionellt krävt stora mängder förkategoriserad områdesspecifik data. På senare år har det tillgängliggjorts förtränade språkmodeller som möjliggjort för språkprocesseringsproblem att lösas med en mindre mängd områdesspecifik kategoriserad data. I den här uppsatsen utforskas datautöknings påverkan på en maskininlärningsmodell för identifiering av namngivna entiteter. De experimentella resultaten indikerar att datautökning förbättrar modellerna, men att inverkan blir mindre när mer kategoriserad data är tillgänglig. Sammanfattningsvis så kan datautökning förbättra modeller för identifiering av namngivna entiteter när få förkategoriserade meningar finns tillgängliga för träning.

SOTA language models have demonstrated remarkable capabilities in tackling NLP tasks they have not been explicitly trained on – given a few demonstrations of the task (few-shot learning), or even none at all (zero-shot learning). The purpose of this Master’s thesis has been to investigate practical aspects and potential applications of zero/few-shot learning in the context of text classification. This includes topics such as combined usage with active learning, automated data labeling, and interpretability. Two different methods for zero/few-shot learning have been investigated, and the results indicate that:  • Active learning can be used to marginally improve few-shot performance, but it seems to be mostly beneficial in settings with very few samples (e.g. less than 10). • Zero-shot learning can be used produce reasonable candidate labels for classes in a dataset, given knowledge of the classification task at hand.  • It is difficult to trust the predictions of zero-shot text classification without access to a validation dataset, but IML methods such as saliency maps could find usage in debugging zero-shot models. / Ledande språkmodeller har uppvisat anmärkningsvärda förmågor i att lösa NLP-problem de inte blivit explicit tränade på – givet några exempel av problemet (few-shot learning), eller till och med inga alls (zero-shot learning). Syftet med det här examensarbetet har varit att undersöka praktiska aspekter och potentiella tillämpningar av zero/few-shot learning inom kontext av textklassificering. Detta inkluderar kombinerad användning med aktiv inlärning, automatiserad datamärkning, och tolkningsbarhet. Två olika metoder för zero/few-shot learning har undersökts, och resultaten indikerar att: • Aktiv inlärning kan användas för att marginellt förbättra textklassificering med few-shot learning, men detta verkar vara mest fördelaktigt i situationer med väldigt få datapunkter (t.ex. mindre än 10). • Zero-shot learning kan användas för att hitta lämpliga etiketter för klasser i ett dataset, givet kunskap om klassifikationsuppgiften av intresse. • Det är svårt att lita på robustheten i textklassificering med zero-shot learning utan tillgång till valideringsdata, men metoder inom tolkningsbar maskininlärning såsom saliency maps skulle kunna användas för att felsöka zero-shot modeller.

Dans ce mémoire, nous étudions la généralisation des réseaux de neurones dans le contexte du méta-apprentissage, en analysant divers propriétés des surface leurs fonctions objectifs. La recherche en apprentissage automatique portant sur les surfaces de fonctions objectifs des réseaux de neurones ayant aidé à comprendre leur généralisation en apprentissage supervisé standard, nous proposons l'étude de telles surfaces dans le but d'approfondir nos connaissances sur la généralisation en méta-apprentissage. Nous introduisons d'abord la littérature sur les fonctions objectifs des réseaux de neurones à la Section \ref{sec:intro:objective_landscapes}, puis celle portant sur le méta-apprentissage à la Section \ref{sec:intro:meta-learning}, pour enfin terminer notre introduction avec le méta-apprentissage par descente de gradient, très similaire à l'entraînement des réseaux de neurones par descente de gradient stochastique et pour une tâche unique. Nous présentons par la suite notre travail sur les fonctions objectifs en méta-apprentissage au Chapitre \ref{chap:prof_forcing}, lequel nous avons soumis à la conférence NeurIPS 2019 en tant qu'article scientifique. Au moment d'écrire ce mémoire, et au meilleur de notre connaissance, ce travail est le premier à étudier empiriquement les surfaces des fonctions objectifs en méta-apprentissage, particulièrement dans le contexte de l'apprentissage profond, et nous mettons notamment en lumière certaines propriétés de ces surfaces qui apparaissent liées à la généralisation des réseaux de neurones à de nouvelles tâches. Nous démontrons empiriquement qu'alors que progresse la phase de méta-entraînement, pour les solutions aux nouvelles tâches obtenues via quelques itérations de descente de gradient, la courbure de la fonction objective décroit monotoniquement, la valeur de la fonction objective diminue, tandis que la distance euclidienne avec la solution ``méta-entraînement" augmente. Cependant, nous observons que la courbure des minima continue de décroître même lorsque le sur-apprentissage devient apparent et que la généralisation commence à se dégrader, indiquant que la courbure des minima semble peu corrélée à la généralisation en méta-apprentissage par descente de gradient. De plus, nous montrons empiriquement que la généralisation aux nouvelles tâches semble plutôt liée à la cohérence de leurs trajectoires d'adaptation dans l'espace des paramètres, mesurée par la similarité cosinus moyenne entre les trajectoires. Nous montrons également que la cohérence des gradients ''meta-test", mesurée par le produit scalaire moyen entre les vecteurs de gradients spécifiques aux nouvelles tâches, évalué à solution meta-entraînement, est également corrélée à la généralisation. Nous basant sur ces observations, nous proposons un nouveau terme de régularisation pour l'algorithme de méta-apprentissage Model Agnostic Meta-Learning (MAML). / In this master's thesis, we study the generalization of neural networks in gradient-based meta-learning by analyzing various properties of the objective landscapes. Meta-learning, a challenging paradigm where models not only have to learn a task but beyond that, are trained for ``learning to learn" as they must adapt to new tasks and environments with very limited data about them. With research on the objective landscapes of neural networks in classical supervised having provided some answers regarding their ability to generalize for new data points, we propose similar analyses aimed at understanding generalization in meta-learning. We first introduce the literature on objective landscapes of neural networks in Section \ref{sec:intro:objective_landscapes}. We then introduce the literature of meta-learning in Section \ref{chap:prof_forcing}, concluding our introduction with the approach of gradient-based meta-learning, a meta-learning setup that bears strong similarities to the traditional supervised learning setup through stochastic gradient-based optimization. At the time of writing of this thesis, and to the best of our knowledge, this is the first work to empirically study the objective landscapes in gradient-based meta-learning, especially in the context of deep learning. We notably provide some insights on some properties of those landscapes that appear correlated to the generalization to new tasks. We experimentally demonstrate that as meta-training progresses, the meta-test solutions, obtained after adapting the meta-train solution of the model, to new tasks via few steps of gradient-based fine-tuning, become flatter, lower in loss, and further away from the meta-train solution. We also show that those meta-test solutions become flatter even as generalization starts to degrade, thus providing experimental evidence against the correlation between generalization and flat minima in the paradigm of gradient-based meta-leaning. Furthermore, we provide empirical evidence that generalization to new tasks is correlated with the coherence between their adaptation trajectories in parameter space, measured by the average cosine similarity between task-specific trajectory directions, starting from a same meta-train solution. We also show that coherence of meta-test gradients, measured by the average inner product between the task-specific gradient vectors evaluated at meta-train solution, is also correlated with generalization. Based on these observations, we propose a novel regularizer for the Model Agnostic Meta-Learning (MAML) algorithm and provide experimental evidence for its effectiveness.

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.

Tesis por compendio / [ES] En la última década, el aprendizaje profundo (DL) se ha convertido en la principal herramienta para las tareas de visión por ordenador (CV). Bajo el paradigma de aprendizaje supervisado, y gracias a la recopilación de grandes conjuntos de datos, el DL ha alcanzado resultados impresionantes utilizando redes neuronales convolucionales (CNNs). Sin embargo, el rendimiento de las CNNs disminuye cuando no se dispone de suficientes datos, lo cual dificulta su uso en aplicaciones de CV en las que sólo se dispone de unas pocas muestras de entrenamiento, o cuando el etiquetado de imágenes es una tarea costosa. Estos escenarios motivan la investigación de estrategias de aprendizaje menos supervisadas. En esta tesis, hemos explorado diferentes paradigmas de aprendizaje menos supervisados. Concretamente, proponemos novedosas estrategias de aprendizaje autosupervisado en la clasificación débilmente supervisada de imágenes histológicas gigapixel. Por otro lado, estudiamos el uso del aprendizaje por contraste en escenarios de aprendizaje de pocos disparos para la vigilancia automática de cruces de ferrocarril. Por último, se estudia la localización de lesiones cerebrales en el contexto de la segmentación no supervisada de anomalías. Asimismo, prestamos especial atención a la incorporación de conocimiento previo durante el entrenamiento que pueda mejorar los resultados en escenarios menos supervisados. En particular, introducimos proporciones de clase en el aprendizaje débilmente supervisado en forma de restricciones de desigualdad. Además, se incorpora la homogeneización de la atención para la localización de anomalías mediante términos de regularización de tamaño y entropía. A lo largo de esta tesis se presentan diferentes métodos menos supervisados de DL para CV, con aportaciones sustanciales que promueven el uso de DL en escenarios con datos limitados. Los resultados obtenidos son prometedores y proporcionan a los investigadores nuevas herramientas que podrían evitar la anotación de cantidades masivas de datos de forma totalmente supervisada. / [CA] En l'última dècada, l'aprenentatge profund (DL) s'ha convertit en la principal eina per a les tasques de visió per ordinador (CV). Sota el paradigma d'aprenentatge supervisat, i gràcies a la recopilació de grans conjunts de dades, el DL ha aconseguit resultats impressionants utilitzant xarxes neuronals convolucionals (CNNs). No obstant això, el rendiment de les CNNs disminueix quan no es disposa de suficients dades, la qual cosa dificulta el seu ús en aplicacions de CV en les quals només es disposa d'unes poques mostres d'entrenament, o quan l'etiquetatge d'imatges és una tasca costosa. Aquests escenaris motiven la investigació d'estratègies d'aprenentatge menys supervisades. En aquesta tesi, hem explorat diferents paradigmes d'aprenentatge menys supervisats. Concretament, proposem noves estratègies d'aprenentatge autosupervisat en la classificació feblement supervisada d'imatges histològiques gigapixel. D'altra banda, estudiem l'ús de l'aprenentatge per contrast en escenaris d'aprenentatge de pocs trets per a la vigilància automàtica d'encreuaments de ferrocarril. Finalment, s'estudia la localització de lesions cerebrals en el context de la segmentació no supervisada d'anomalies. Així mateix, prestem especial atenció a la incorporació de coneixement previ durant l'entrenament que puga millorar els resultats en escenaris menys supervisats. En particular, introduïm proporcions de classe en l'aprenentatge feblement supervisat en forma de restriccions de desigualtat. A més, s'incorpora l'homogeneïtzació de l'atenció per a la localització d'anomalies mitjançant termes de regularització de grandària i entropia. Al llarg d'aquesta tesi es presenten diferents mètodes menys supervisats de DL per a CV, amb aportacions substancials que promouen l'ús de DL en escenaris amb dades limitades. Els resultats obtinguts són prometedors i proporcionen als investigadors noves eines que podrien evitar l'anotació de quantitats massives de dades de forma totalment supervisada. / [EN] In the last decade, deep learning (DL) has become the main tool for computer vision (CV) tasks. Under the standard supervised learnng paradigm, and thanks to the progressive collection of large datasets, DL has reached impressive results on different CV applications using convolutional neural networks (CNNs). Nevertheless, CNNs performance drops when sufficient data is unavailable, which creates challenging scenarios in CV applications where only few training samples are available, or when labeling images is a costly task, that require expert knowledge. Those scenarios motivate the research of not-so-supervised learning strategies to develop DL solutions on CV. In this thesis, we have explored different less-supervised learning paradigms on different applications. Concretely, we first propose novel self-supervised learning strategies on weakly supervised classification of gigapixel histology images. Then, we study the use of contrastive learning on few-shot learning scenarios for automatic railway crossing surveying. Finally, brain lesion segmentation is studied in the context of unsupervised anomaly segmentation, using only healthy samples during training. Along this thesis, we pay special attention to the incorporation of tasks-specific prior knowledge during model training, which may be easily obtained, but which can substantially improve the results in less-supervised scenarios. In particular, we introduce relative class proportions in weakly supervised learning in the form of inequality constraints. Also, attention homogenization in VAEs for anomaly localization is incorporated using size and entropy regularization terms, to make the CNN to focus on all patterns for normal samples. The different methods are compared, when possible, with their supervised counterparts. In short, different not-so-supervised DL methods for CV are presented along this thesis, with substantial contributions that promote the use of DL in data-limited scenarios. The obtained results are promising, and provide researchers with new tools that could avoid annotating massive amounts of data in a fully supervised manner. / The work of Julio Silva Rodríguez to carry out this research and to elaborate this dissertation has been supported by the Spanish Government under the FPI Grant PRE2018-083443. / Silva Rodríguez, JJ. (2022). Learning from limited labelled data: contributions to weak, few-shot, and unsupervised learning [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/190633 / Compendio

In recent years, there has been a significant increase in the usage of audience engagement platforms, which have allowed for engaging interactions between presenters and their audiences. The increased popularity of the platforms comes from the fact that engaging and interactive presentations have been shown to improve learning outcomes and create positive presentation experiences. However, using the platforms does not guarantee that your audience is engaged and participating. Given that the added value of engaging presentations only applies if the audience is actually engaged, it increases the need to know if and how engaged your audience is. The usage of audience engagement platforms has allowed for new ways of engagement to be studied. By utilizing the data gathered from the interactive presentation sessions, engagement can be studied and quantified through the modeling of the data. As the usage of audience engagement platforms and the study of presentation engagement is relatively new, there exists a limited amount of labeled data quantifying the level of engagement during presentations. To model the data, machine learning models should therefore be trained to generalize by being exposed to a limited number of presentation samples. This technique of training machine learning models is also referred to as few-shot learning. Distance-based machine learning models are defined in this study as models that make classifications and inferences by calculating distances between observations or observation class representations. Distance-based models have previously shown relatively good performance in few-shot learning applications, and interest therefore lies in expanding their application areas. This study presents a comparative evaluation of distance- and non-distance-based machine learning models given the problem of classifying presentations as being engaged or non-engaged, and estimating presentation class likelihoods in a few-shot learning context. A presentation-level dataset was gathered from the interactive presentation sessions, and each presentation observation was labeled as being engaged or non-engaged. The machine learning models were then trained to model the data and evaluated in terms of how well they were able to generalize to unseen testing samples by being exposed to a limited number of training observations. In particular, their classification and class likelihood estimation performances were evaluated. The results conclude that the distance-based models outperformed the non-distance-based models artificial neural network and relevance-vector machine given the presentation class likelihood estimation problem. The metric learning nearest neighbor classifier was the only distance-based model that outperformed all the non-distance-based models given both the presentation classification and class likelihood estimation problems. / Under de senaste åren har det skett en betydande ökning av användningen av plattformar för publikengagemang, vilket har möjliggjort engagerande interaktioner mellan presentatörer och deras publik. Plattformarnas ökade popularitet kommer från det faktum att engagerande och interaktiva presentationer har visat sig förbättra läranderesultat och skapa positiva presentationsupplevelser. Att använda plattformarna garanterar dock inte att din publik är engagerad och deltagande. Med tanke på att mervärdet av engagerande presentationer bara gäller om publiken faktiskt är engagerad, ökar det behovet av att veta om och hur engagerad din publik är. Användningen av plattformar för publikengagemang har gjort det möjligt att på nya sätt studera engagemang. Genom att använda data som samlats in från de interaktiva presentationssessionerna kan engagemang studeras och kvantifieras genom modellering av data. Eftersom användandet av plattformar för publikengagemang och studien av presentationsengagemang är relativt nytt, finns det en begränsad mängd märkt data som kvantifierar nivån av engagemang under presentationerna. För att modellera datan så bör maskininlärningsmodeller tränas att generalisera genom att utsättas för ett begränsad antal presentations observationer. Denna teknik för att träna inlärningsmodeller kallas också few-shot lärande. Distans-baserade maskininlärningsmodeller definieras i denna studie som modeller som gör klassificeringar genom att beräkna avstånd mellan observationer eller observationsklass representationer. Distans-baserade modeller har tidigare visat relativt goda resultat i few-shot inlärning problem, och intresset ligger därför i att utöka deras tillämpningsområden. Denna studie presenterar en jämförande utvärdering av distans- och icke-distans baserade maskininlärningsmodeller givet problemet med att klassificera presentationer som engagerade eller icke-engagerade, och uppskattning av presentation klasssannolikheter i ett few-shot inlärnings sammanhang. Ett dataset på presentationsnivå samlades in från de interaktiva presentationssessionerna, och varje presentation märktes som engagerad eller icke-engagerad. Maskininlärningsmodellerna tränades sedan för att modellera data och utvärderades i termer av hur väl de kunde generalisera till osedda testobservationer givet att de exponeras mot ett begränsat antal träningsobservationer. I synnerhet utvärderades deras klassificering och uppskattning av klasssannolikheter. Resultaten visade att alla distans-baserade modeller var bättre än de icke-distansbaserade modellerna artificial neural network och relevence-vector machine givet problemet med uppskattning av klasssannolikheter. Den distans-baserade metric learning nearest neighbor klassificeraren var den enda avståndsbaserade modellen som överträffade alla icke-distansbaserade modeller givet problemen med presentations klassificering och klasssannolikhets uppskattning.

<p>Nuclear engineering hosts diverse domains including, but not limited to, power plant automation, human-machine interfacing, detection and identification of special nuclear materials, modeling of reactor kinetics and dynamics that most frequently are described by systems of differential equations (DEs), either ordinary (ODEs) or partial ones (PDEs). In this work we study multiple problems related to safety and Special Nuclear Material detection, and numerical solutions for partial differential equations using neural networks. More specifically, this work is divided in six chapters. Chapter 1 is the introduction, in Chapter</p> <p>2 we discuss the development of a gamma-ray radionuclide library for the characterization</p> <p>of gamma-spectra. In Chapter 3, we present a new approach, the ”Variance Counterbalancing”, for stochastic</p> <p>large-scale learning. In Chapter 4, we introduce a systematic approach for constructing proper trial solutions to partial differential equations (PDEs) of up to second order, using neural forms that satisfy prescribed initial, boundary and interface conditions. Chapter 5 is about an alternative, less imposing development of neural-form trial solutions for PDEs, inside rectangular and non-rectangular convex boundaries. Chapter 6 presents an ensemble method that avoids the multicollinearity issue and provides</p> <p>enhanced generalization performance that could be suitable for handling ”few-shots”- problems frequently appearing in nuclear engineering.</p>

Les modélisateurs rencontrent souvent des défis ou des difficultés lorsqu’il s’agit de concevoir un modèle logiciel particulier. Dans cette thèse, nous avons exploré différentes voies et examiné différentes approches pour résoudre cette problématique. Nous proposons enfin une approche simple mais novatrice qui améliore la complétion des activités de modélisation de domaines. Cette approche exploite la puissance des modèles de langage de grande taille en utilisant l’apprentissage par seulement quelques exemples, éliminant ainsi la nécessité d’un apprentissage profond ou d’un ajustement fin (fine tuning) sur des ensembles de données rares dans ce domaine. L’un des points forts de notre approche est sa polyvalence, car elle peut s’intégrer fa cilement à de nombreuses activités de modélisation, fournissant un aide précieux et des recommendations aux modélisateurs. De plus, nous avons mené une étude utilisateur pour évaluer l’utilité de cette méthode et la valeur de l’assistance en modélisation; nous avons cherché à savoir si l’effort investi dans l’assistance en modélisation vaut la peine en recueillant les commentaires des concepteurs de modèles logiciels. / Modelers often encounter challenges or difficulties when it comes to designing a particular software model. Throughout this thesis, we have explored various paths and examined different approaches to address this issue. We finally propose a simple yet novel approach enhancing completion in domain modeling activities. This approach leverages the power of large language models by utilizing few-shot prompt learning, eliminating the need for extensive training or fine-tuning on scarce datasets in this field. One of the notable strengths of our approach lies in its versatility, as it can be seamlessly integrated into various modeling activities, providing valuable support and recommendations to software modelers. Additionally, we conducted a user study to evaluate the usefulness of this approach and determine the value of providing assistance in modeling; we aimed to determine if the effort invested in modeling assistance is worthwhile by gathering feedback from software modelers.

L'intelligence artificielle est un domaine en pleine évolution. Au premier plan des percées récentes se retrouve des approches connues sous le nom d'apprentissage automatique. Cependant, bien que l'apprentissage automatique ait montré des performances remarquables dans des tâches telles que la reconnaissance et la génération d'images, la génération et la traduction de textes et le traitement de la parole, il est connu pour échouer silencieusement dans des conditions courantes. Cela est dû au fait que les algorithmes modernes héritent des biais des données utilisées pour les créer, ce qui conduit à des prédictions incorrectes lorsqu'ils rencontrent de nouvelles données différentes des données d'entraînement. Ce problème est connu sous le nom de défaillance hors-distribution. Cela rend l'intelligence artificielle moderne peu fiable et constitue un obstacle important à son déploiement sécuritaire et généralisé. Ignorer l'échec de généralisation hors-distribution de l'apprentissage automatique pourrait entraîner des situations mettant des vies en danger. Cette thèse vise à aborder cette question et propose des solutions pour assurer le déploiement sûr et fiable de modèles d'intelligence artificielle modernes. Nous présentons trois articles qui couvrent différentes directions pour résoudre l'échec de généralisation hors-distribution de l'apprentissage automatique. Le premier article propose une approche directe qui démontre une performance améliorée par rapport à l'état de l'art. Le deuxième article établie les bases de recherches futures en généralisation hors distribution dans les séries temporelles, tandis que le troisième article fournit une solution simple pour corriger les échecs de généralisation des grands modèles pré-entraînés lorsqu'entraîné sur tes tâches en aval. Ces articles apportent des contributions précieuses au domaine et fournissent des pistes prometteuses pour la recherche future en généralisation hors distribution. / Artificial Intelligence (AI) is a rapidly advancing field, with data-driven approaches known as machine learning, at the forefront of many recent breakthroughs. However, while machine learning have shown remarkable performance in tasks such as image recognition and generation, text generation and translation, and speech processing, they are known to silently fail under common conditions. This is because modern AI algorithms inherit biases from the data used to train them, leading to incorrect predictions when encountering new data that is different from the training data. This problem is known as distribution shift or out-of-distribution (OOD) failure. This causes modern AI to be untrustworthy and is a significant barrier to the safe widespread deployment of AI. Failing to address the OOD generalization failure of machine learning could result in situations that put lives in danger or make it impossible to deploy AI in any significant manner. This thesis aims to tackle this issue and proposes solutions to ensure the safe and reliable deployment of modern deep learning models. We present three papers that cover different directions in solving the OOD generalization failure of machine learning. The first paper proposes a direct approach that demonstrates improved performance over the state-of-the-art. The second paper lays the groundwork for future research in OOD generalization in time series, while the third paper provides a straightforward solution for fixing generalization failures of large pretrained models when finetuned on downstream tasks. These papers make valuable contributions to the field and provide promising avenues for future research in OOD generalization.