Recent years have witnessed tremendous success in using deep learning approaches to handle low-level vision problems. Most of the deep learning based methods address the low-level vision problem by training a neural network to approximate the mapping from the inputs to the desired ground truths. However, directly learning this mapping is usually difficult and cannot achieve ideal performance. Besides, under the setting of unsupervised learning, the general deep learning approach cannot be used. In this thesis, we investigate and address several problems in low-level vision using the proposed approaches. To learn a better mapping using the existing data, an indirect domain shift mechanism is proposed to add explicit constraints inside the neural network for single image dehazing. This allows the neural network to be optimized across several identified neighbours, resulting in a better performance. Despite the success of the proposed approaches in learning an improved mapping from the inputs to the targets, three problems of unsupervised learning is also investigated. For unsupervised monocular depth estimation, a teacher-student network is introduced to strategically integrate both supervised and unsupervised learning benefits. The teacher network is formed by learning under the binocular depth estimation setting, and the student network is constructed as the primary network for monocular depth estimation. In observing that the performance of the teacher network is far better than that of the student network, a knowledge distillation approach is proposed to help improve the mapping learned by the student. For single image dehazing, the current network cannot handle different types of haze patterns as it is trained on a particular dataset. The problem is formulated as a multi-domain dehazing problem. To address this issue, a test-time training approach is proposed to leverage a helper network in assisting the dehazing network adapting to a particular domain using self-supervision. In lossy compression system, the target distribution can be different from that of the source and ground truths are not available for reference. Thus, the objective is to transform the source to target under a rate constraint, which generalizes the optimal transport. To address this problem, theoretical analyses on the trade-off between compression rate and minimal achievable distortion are studied under the cases with and without common randomness. A deep learning approach is also developed using our theoretical results for addressing super-resolution and denoising tasks. Extensive experiments and analyses have been conducted to prove the effectiveness of the proposed deep learning based methods in handling the problems in low-level vision. / Thesis / Doctor of Philosophy (PhD)

Dynamic Spectrum Access (DSA) has strong potential to address the need for improved spectrum efficiency. Unfortunately, traditional DSA approaches such as simple "sense-and-avoid" fail to provide sufficient performance in many scenarios. Thus, the combination of sensing with deep reinforcement learning (DRL) has been shown to be a promising alternative to previously proposed simplistic approaches. DRL does not require the explicit estimation of transition probability matrices and prohibitively large matrix computations as compared to traditional reinforcement learning methods. Further, since many learning approaches cannot solve the resulting online Partially-Observable Markov Decision Process (POMDP), Deep Recurrent Q-Networks (DRQN) have been proposed to determine the optimal channel access policy via online learning. The fundamental goal of this dissertation is to develop DRL-based solutions to address this POMDP-DSA problem. We mainly consider three aspects in this work: (1) optimal transmission strategies, (2) combined intelligent sensing and transmission strategies, and (c) learning efficiency or online convergence speed. Four key challenges in this problem are (1) the proposed DRQN-based node does not know the other nodes' behavior patterns a priori and must to predict the future channel state based on previous observations; (2) the impact to primary user throughput during learning and even after learning must be limited; (3) resources can be wasted the sensing/observation; and (4) convergence speed must be improved without impacting performance performance. We demonstrate in this dissertation, that the proposed DRQN can learn: (1) the optimal transmission strategy in a variety of environments under partial observations; (2) a sensing strategy that provides near-optimal throughput in different environments while dramatically reducing the needed sensing resources; (3) robustness to imperfect observations; (4) a sufficiently flexible approach that can accommodate dynamic environments, multi-channel transmission and the presence of multiple agents; (5) in an accelerated fashion utilizing one of three different approaches. / Doctor of Philosophy / With the development of wireless communication, such as 5G, global mobile data traffic has experienced tremendous growth, which makes spectrum resources even more critical for future networks. However, the spectrum is an exorbitant and scarce resource. Dynamic Spectrum Access (DSA) has strong potential to address the need for improved spectrum efficiency. Unfortunately, traditional DSA approaches such as simple "sense-and-avoid" fail to provide sufficient performance in many scenarios. Thus, the combination of sensing with deep reinforcement learning (DRL) has been shown to be a promising alternative to previously proposed simplistic approaches. Compared with traditional reinforcement learning methods, DRL does not require explicit estimation of transition probability matrices and extensive matrix computations. Furthermore, since many learning methods cannot solve the resulting online partially observable Markov decision process (POMDP), a deep recurrent Q-network (DRQN) is proposed to determine the optimal channel access policy through online learning. The basic goal of this paper is to develop a DRL-based solution to this POMDP-DSA problem. This paper mainly focuses on improving performance from three directions. 1. Find the optimal (or sub-optimal) channel access strategy based on fixed partial observation mode; 2. Based on work 1, propose a more intelligent way to dynamically and efficiently find more reasonable (higher efficiency) sensing/observation policy and corresponding channel access strategy; 3. On the premise of ensuring performance, use different machine learning algorithms or structures to improve learning efficiency and avoid users waiting too long for expected performance. Through the research in these three main directions, we have found an efficient and diverse solution, namely DRQN-based technology.

Defect prediction models assist tester practitioners on prioritizing the most defect-prone parts of the software. The approach called Cross-Project Defect Prediction (CPDP) refers to the use of known external projects to compose the training set. This approach is useful when the amount of historical defect data of a company to compose the training set is inappropriate or insufficient. Although the principle is attractive, the predictive performance is a limiting factor. In recent years, several methods were proposed aiming at improving the predictive performance of CPDP models. However, to the best of our knowledge, there is no evidence of which CPDP methods typically perform best. Moreover, there is no evidence on which CPDP methods perform better for a specific application domain. In fact, there is no machine learning algorithm suitable for all domains. The decision task of selecting an appropriate algorithm for a given application domain is investigated in the meta-learning literature. A meta-learning model is characterized by its capacity of learning from previous experiences and adapting its inductive bias dynamically according to the target domain. In this work, we investigate the feasibility of using meta-learning for the recommendation of CPDP methods. In this thesis, three main goals were pursued. First, we provide an experimental analysis to investigate the feasibility of using Feature Selection (FS) methods as an internal procedure to improve the performance of two specific CPDP methods. Second, we investigate which CPDP methods present typically best performances. We also investigate whether the typically best methods perform best for the same project datasets. The results reveal that the most suitable CPDP method for a project can vary according to the project characteristics, which leads to the third investigation of this work. We investigate the several particularities inherent to the CPDP context and propose a meta-learning solution able to learn from previous experiences and recommend a suitable CDPD method according to the characteristics of the project being predicted. We evaluate the learning capacity of the proposed solution and its performance in relation to the typically best CPDP methods. / Modelos de predição de defeitos auxiliam profissionais de teste na priorização de partes do software mais propensas a conter defeitos. A abordagem de predição de defeitos cruzada entre projetos (CPDP) refere-se à utilização de projetos externos já conhecidos para compor o conjunto de treinamento. Essa abordagem é útil quando a quantidade de dados históricos de defeitos é inapropriada ou insuficiente para compor o conjunto de treinamento. Embora o princípio seja atrativo, o desempenho de predição é um fator limitante nessa abordagem. Nos últimos anos, vários métodos foram propostos com o intuito de melhorar o desempenho de predição de modelos CPDP. Contudo, na literatura, existe uma carência de estudos comparativos que apontam quais métodos CPDP apresentam melhores desempenhos. Além disso, não há evidências sobre quais métodos CPDP apresentam melhor desempenho para um domínio de aplicação específico. De fato, não existe um algoritmo de aprendizado de máquina que seja apropriado para todos os domínios de aplicação. A tarefa de decisão sobre qual algoritmo é mais adequado a um determinado domínio de aplicação é investigado na literatura de meta-aprendizado. Um modelo de meta-aprendizado é caracterizado pela sua capacidade de aprender a partir de experiências anteriores e adaptar seu viés de indução dinamicamente de acordo com o domínio alvo. Neste trabalho, nós investigamos a viabilidade de usar meta-aprendizado para a recomendação de métodos CPDP. Nesta tese são almejados três principais objetivos. Primeiro, é conduzida uma análise experimental para investigar a viabilidade de usar métodos de seleção de atributos como procedimento interno de dois métodos CPDP, com o intuito de melhorar o desempenho de predição. Segundo, são investigados quais métodos CPDP apresentam um melhor desempenho em um contexto geral. Nesse contexto, também é investigado se os métodos com melhor desempenho geral apresentam melhor desempenho para os mesmos conjuntos de dados (ou projetos de software). Os resultados revelam que os métodos CPDP mais adequados para um projeto podem variar de acordo com as características do projeto sendo predito. Essa constatação conduz à terceira investigação realizada neste trabalho. Foram investigadas as várias particularidades inerentes ao contexto CPDP a fim de propor uma solução de meta-aprendizado capaz de aprender com experiências anteriores e recomendar métodos CPDP adequados, de acordo com as características do software. Foram avaliados a capacidade de meta-aprendizado da solução proposta e a sua performance em relação aos métodos base que apresentaram melhor desempenho geral.

Defect prediction models assist tester practitioners on prioritizing the most defect-prone parts of the software. The approach called Cross-Project Defect Prediction (CPDP) refers to the use of known external projects to compose the training set. This approach is useful when the amount of historical defect data of a company to compose the training set is inappropriate or insufficient. Although the principle is attractive, the predictive performance is a limiting factor. In recent years, several methods were proposed aiming at improving the predictive performance of CPDP models. However, to the best of our knowledge, there is no evidence of which CPDP methods typically perform best. Moreover, there is no evidence on which CPDP methods perform better for a specific application domain. In fact, there is no machine learning algorithm suitable for all domains. The decision task of selecting an appropriate algorithm for a given application domain is investigated in the meta-learning literature. A meta-learning model is characterized by its capacity of learning from previous experiences and adapting its inductive bias dynamically according to the target domain. In this work, we investigate the feasibility of using meta-learning for the recommendation of CPDP methods. In this thesis, three main goals were pursued. First, we provide an experimental analysis to investigate the feasibility of using Feature Selection (FS) methods as an internal procedure to improve the performance of two specific CPDP methods. Second, we investigate which CPDP methods present typically best performances. We also investigate whether the typically best methods perform best for the same project datasets. The results reveal that the most suitable CPDP method for a project can vary according to the project characteristics, which leads to the third investigation of this work. We investigate the several particularities inherent to the CPDP context and propose a meta-learning solution able to learn from previous experiences and recommend a suitable CDPD method according to the characteristics of the project being predicted. We evaluate the learning capacity of the proposed solution and its performance in relation to the typically best CPDP methods. / Modelos de predição de defeitos auxiliam profissionais de teste na priorização de partes do software mais propensas a conter defeitos. A abordagem de predição de defeitos cruzada entre projetos (CPDP) refere-se à utilização de projetos externos já conhecidos para compor o conjunto de treinamento. Essa abordagem é útil quando a quantidade de dados históricos de defeitos é inapropriada ou insuficiente para compor o conjunto de treinamento. Embora o princípio seja atrativo, o desempenho de predição é um fator limitante nessa abordagem. Nos últimos anos, vários métodos foram propostos com o intuito de melhorar o desempenho de predição de modelos CPDP. Contudo, na literatura, existe uma carência de estudos comparativos que apontam quais métodos CPDP apresentam melhores desempenhos. Além disso, não há evidências sobre quais métodos CPDP apresentam melhor desempenho para um domínio de aplicação específico. De fato, não existe um algoritmo de aprendizado de máquina que seja apropriado para todos os domínios de aplicação. A tarefa de decisão sobre qual algoritmo é mais adequado a um determinado domínio de aplicação é investigado na literatura de meta-aprendizado. Um modelo de meta-aprendizado é caracterizado pela sua capacidade de aprender a partir de experiências anteriores e adaptar seu viés de indução dinamicamente de acordo com o domínio alvo. Neste trabalho, nós investigamos a viabilidade de usar meta-aprendizado para a recomendação de métodos CPDP. Nesta tese são almejados três principais objetivos. Primeiro, é conduzida uma análise experimental para investigar a viabilidade de usar métodos de seleção de atributos como procedimento interno de dois métodos CPDP, com o intuito de melhorar o desempenho de predição. Segundo, são investigados quais métodos CPDP apresentam um melhor desempenho em um contexto geral. Nesse contexto, também é investigado se os métodos com melhor desempenho geral apresentam melhor desempenho para os mesmos conjuntos de dados (ou projetos de software). Os resultados revelam que os métodos CPDP mais adequados para um projeto podem variar de acordo com as características do projeto sendo predito. Essa constatação conduz à terceira investigação realizada neste trabalho. Foram investigadas as várias particularidades inerentes ao contexto CPDP a fim de propor uma solução de meta-aprendizado capaz de aprender com experiências anteriores e recomendar métodos CPDP adequados, de acordo com as características do software. Foram avaliados a capacidade de meta-aprendizado da solução proposta e a sua performance em relação aos métodos base que apresentaram melhor desempenho geral.

Notre thèse de pharmacie clinique est une contribution à l’optimisation de l’efficience du traitement antirétroviral (TAR) par des méthodes de modélisation en Côte d’Ivoire. La première étude a été consacrée à une modélisation de la réponse antirétrovirale par taxonomie des trajectoires de taux de CD4 en utilisant un modèle de méta-apprentissage des trajectoires d’indicateurs biomédicaux. Ce modèle appliqué à la taxonomie des trajectoires des taux de CD4 a montré son intérêt dans la mise en évidence de classes distinctes de patients avec des caractéristiques particulières justifiant et/ou déterminant le profil particulier de méta-trajectoires de leur marqueur immunologique au cours du traitement. La deuxième tâche a consisté en une évaluation de l’impact de principaux déterminants des méta-trajectoires de taux de CD4 sur divers types de réponse immunologique à partir d’un modèle explicatif avec une équation de régression logistique. Les réponses immunologiques considérées ont été exprimées en termes d’absence de gain de CD4, de gain sub-optimal et de gain optimal de CD4 à différentes périodes de suivi du TAR. Enfin l’évaluation de l’efficience des stratégies antirétrovirales de première ligne en Côte d’Ivoire, a été abordée dans la dernière partie avec un modèle pharmaco-économique. Nous avons effectué une étude préliminaire ouvrant des perspectives pour encourager la mise en oeuvre d’évaluations pharmaco-économiques complètes par modélisation en Côte d’Ivoire. Elle a permis de mettre en exergue les parties méthodologiques pouvant être sujettes à caution dans une étude de modélisation pharmaco-économique des TAR de première ligne dans un contexte de ressources limitées / Our thesis of Clinical pharmacy is a contribution to optimize the efficiency of antiretroviral therapy (ART) by modeling methods in Côte d'Ivoire. The first study was devoted to modeling the antiretroviral response from taxonomy of CD4 counts trajectories, using a meta-learning model of biomedical markers trajectories. This model applied to the taxonomy of the CD4 counts trajectories showed its interest in the identification of distinct classes of patients with particular characteristics justifying and/or determining the specific profile of meta-trajectories of the immunological marker during treatment. The second task was an assessment of the impact of key determinants of CD4 counts meta-trajectories on various types of immune response from an explanatory model with a logistic regression equation. Antiretroviral immune responses considered were expressed in terms of absence of CD4 gain, sub-optimal gain and optimal gain of CD4 at different periods of follow-up of ART. Finally the evaluation of the efficiency of first-line antiretroviral strategies in Côte d'Ivoire, was discussed in the last part with a projective pharmaco-economic model. We conducted a preliminary exploratory study opening up prospects to encourage the implementation of comprehensive pharmaco-economic assessments by modeling in Côte d'Ivoire. This study helped to highlight the unreliable methodological sections in a pharmaco-economic modeling of first-line ART in resource-limited settings

Le problème “d’apprentissage continu” implique l’entraînement des modèles profonds avec une capacité limitée qui doivent bien fonctionner sur un nombre inconnu de tâches arrivant séquentiellement. Cette configuration peut souvent résulter en un système d’apprentissage qui souffre de “l’oublie catastrophique”, lorsque l’apprentissage d’une nouvelle tâche provoque des interférences sur la progression de l’apprentissage des anciennes tâches. Les travaux récents ont montré que les techniques de “méta-apprentissage” ont le potentiel de ré- duire les interférences entre les anciennes et les nouvelles tâches. Cependant, les procé- dures d’entraînement ont présentement une tendance à être lente ou hors ligne et sensibles à de nombreux hyperparamètres. Dans ce travail, nous proposons “Look-ahead MAML (La-MAML)”, un algorithme de méta-apprentissage rapide basé sur l’optimisation pour l’apprentissage continu en ligne et aidé par une petite mémoire épisodique. Ceci est réalisé en utilisant l’équivalence d’un objectif MAML en plusieurs étapes et un objectif d’apprentissage continu “temps conscient”. L’équivalence résulte au développement d’un algorithme intuitif que nous appelons Continual-MAML (C-MAML), utilisant un méta-apprentissage continu pour optimiser un modèle afin qu’il fonctionne bien sur une série de distributions de don- nées changeantes. En intégrant la modulation des taux d’apprentissage par paramètre dans La-MAML, notre approche fournit un moyen plus flexible et efficace d’atténuer l’oubli catas- trophique par rapport aux méthodes classiques basées sur les prieurs. Cette modulation a également des liens avec des travaux sur la métadescendance, que nous identifions comme une direction importante de la recherche pour développer de meilleurs optimiser pour un ap- prentissage continu. Dans des expériences menées sur des repères de classification visuelle du monde réel, La-MAML atteint des performances supérieures aux autres approches basées sur la relecture, basées sur les prieurs et basées sur le méta-apprentissage pour un apprentissage continu. Nous démontrons également qu’elle est robuste et plus évolutive que de nombreuses approches de pointe. / The continual learning problem involves training models with limited capacity to perform well on a set of an unknown number of sequentially arriving tasks. This setup can of- ten see a learning system undergo catastrophic forgetting, when learning a newly seen task causes interference on the learning progress of old tasks. While recent work has shown that meta-learning has the potential to reduce interference between old and new tasks, the current training procedures tend to be either slow or offline, and sensitive to many hyper-parameters. In this work, we propose Look-ahead MAML (La-MAML), a fast optimisation-based meta- learning algorithm for online-continual learning, aided by a small episodic memory. This is achieved by realising the equivalence of a multi-step MAML objective to a time-aware con- tinual learning objective adopted in prior work. The equivalence leads to the formulation of an intuitive algorithm that we call Continual-MAML (C-MAML), employing continual meta- learning to optimise a model to perform well across a series of changing data distributions. By additionally incorporating the modulation of per-parameter learning rates in La-MAML, our approach provides a more flexible and efficient way to mitigate catastrophic forgetting compared to conventional prior-based methods. This modulation also has connections to prior work on meta-descent, which we identify as an important direction of research to de- velop better optimizers for continual learning. In experiments conducted on real-world visual classification benchmarks, La-MAML achieves performance superior to other replay-based, prior-based and meta-learning based approaches for continual learning. We also demonstrate that it is robust, and more scalable than many recent state-of-the-art approaches.

Surrogate-model-based optimization is widely used to solve black-box optimization problems if the evaluation of a target system is expensive. However, when the optimization budget is limited to a single or several evaluations, surrogate-model-based optimization may not perform well due to the lack of knowledge about the search space. In this case, transfer learning helps to get a good optimization result due to the usage of experience from the previous optimization runs. And if the budget is not strictly limited, transfer learning is capable of improving the final results of black-box optimization. The recent work in surrogate-model-based optimization showed that using multiple surrogates (i.e., applying multi-surrogate-model optimization) can be extremely efficient in complex search spaces. The main assumption of this thesis suggests that transfer learning can further improve the quality of multi-surrogate-model optimization. However, to the best of our knowledge, there exist no approaches to transfer learning in the multi-surrogate-model context yet. In this thesis, we propose an approach to transfer learning for multi-surrogate-model optimization. It encompasses an improved method of defining the expediency of knowledge transfer, adapted multi-surrogate-model recommendation, multi-task learning parameter tuning, and few-shot learning techniques. We evaluated the proposed approach with a set of algorithm selection and parameter setting problems, comprising mathematical functions optimization and the traveling salesman problem, as well as random forest hyperparameter tuning over OpenML datasets. The evaluation shows that the proposed approach helps to improve the quality delivered by multi-surrogate-model optimization and ensures getting good optimization results even under a strictly limited budget.:1 Introduction 1.1 Motivation 1.2 Research objective 1.3 Solution overview 1.4 Thesis structure 2 Background 2.1 Optimization problems 2.2 From single- to multi-surrogate-model optimization 2.2.1 Classical surrogate-model-based optimization 2.2.2 The purpose of multi-surrogate-model optimization 2.2.3 BRISE 2.5.0: Multi-surrogate-model-based software product line for parameter tuning 2.3 Transfer learning 2.3.1 Definition and purpose of transfer learning 2.4 Summary of the Background 3 Related work 3.1 Questions to transfer learning 3.2 When to transfer: Existing approaches to determining the expediency of knowledge transfer 3.2.1 Meta-features-based approaches 3.2.2 Surrogate-model-based similarity 3.2.3 Relative landmarks-based approaches 3.2.4 Sampling landmarks-based approaches 3.2.5 Similarity threshold problem 3.3 What to transfer: Existing approaches to knowledge transfer 3.3.1 Ensemble learning 3.3.2 Search space pruning 3.3.3 Multi-task learning 3.3.4 Surrogate model recommendation 3.3.5 Few-shot learning 3.3.6 Other approaches to transferring knowledge 3.4 How to transfer (discussion): Peculiarities and required design decisions for the TL implementation in multi-surrogate-model setup 3.4.1 Peculiarities of model recommendation in multi-surrogate-model setup 3.4.2 Required design decisions in multi-task learning 3.4.3 Few-shot learning problem 3.5 Summary of the related work analysis 4 Transfer learning for multi-surrogate-model optimization 4.1 Expediency of knowledge transfer 4.1.1 Experiments’ similarity definition as a variability point 4.1.2 Clustering to filter the most suitable experiments 4.2 Dynamic model recommendation in multi-surrogate-model setup 4.2.1 Variable recommendation granularity 4.2.2 Model recommendation by time and performance criteria 4.3 Multi-task learning 4.4 Implementation of the proposed concept 4.5 Conclusion of the proposed concept 5 Evaluation 5.1 Benchmark suite 5.1.1 APSP for the meta-heuristics 5.1.2 Hyperparameter optimization of the Random Forest algorithm 5.2 Environment setup 5.3 Evaluation plan 5.4 Baseline evaluation 5.5 Meta-tuning for a multi-task learning approach 5.5.1 Revealing the dependencies between the parameters of multi-task learning and its performance 5.5.2 Multi-task learning performance with the best found parameters 5.6 Expediency determination approach 5.6.1 Expediency determination as a variability point 5.6.2 Flexible number of the most similar experiments with the help of clustering 5.6.3 Influence of the number of initial samples on the quality of expediency determination 5.7 Multi-surrogate-model recommendation 5.8 Few-shot learning 5.8.1 Transfer of the built surrogate models’ combination 5.8.2 Transfer of the best configuration 5.8.3 Transfer from different experiment instances 5.9 Summary of the evaluation results 6 Conclusion and Future work

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.

No description available.

Using automatic methods to assign incoming tracks and albums from multiple sources to artists entities in a digital rights management company, where no universal artist identifier is available and artist names can be ambiguous, is a challenging problem. In this work we propose to use stacked generalization to combine the predictions of heterogeneous classifiers for an improved quality of artist content matching on two datasets from a digital rights management company. We compare the performance of using a nonlinear meta-level learner to a linear meta-level learner for the stacked generalization on the two datasets, as well as on eight additional datasets to see how well our results general- ize. We conduct experiments and evaluate how the different meta-level learners perform, using the base learners’ class probabilities or a combination of the base learners’ class probabilities and original input features as meta-features. Our results indicate that stacking with a non-linear meta-level learner can improve predictions on the artist chooser problem. Furthermore, our results indicate that when using a linear meta-level learner for stacked generalization, using the base learners’ class probabilities as metafeatures works best, while using a combination of the base learners’ class probabilities and the original input features as meta-features works best when using a non-linear metalevel learner. Among all the evaluated stacking approaches, stacking with a non-linear meta-level learner, using a combination of the base learners’ class probabilities and the original input features as meta-features, performs the best in our experiments over the ten evaluation datasets. / Att använda automatiska metoder för att tilldela spår och album från olika källor till artister i en digital underhållningstjänst är problematiskt då det inte finns några universellt använda identifierare för artister och namn på artister kan vara tvetydiga. I det här verket föreslår vi en användning av staplad generalisering för att kombinera förutsägningar från heterogena klassificerare för förbättra artistmatchningen i två datamäng från en digital underhållningstjänst. Vi jämför prestandan mellan en linjär och en icke-linjär metainlärningsmetod för den staplade generaliseringen av de två datamängder, samt även åtta ytterligare datamäng för att se hur resultaten kan generaliseras. Vi utför experiment och utvärderar hur de olika metainlärningsmetoderna presterar genom att använda basinlärningsmetodens klassannolikheter eller en kombination av basinlärningsmetodens klassannolikheter och den ursprungliga representationen som metarepresentation. Våra resultat indikerar att staplandet med en icke-linjär metainlärningsmetod kan förbättra förutsägningarna i problemet med att tilldela artister. Vidare indikerar våra resultat att när man använder en linjär metainlärningsmetod för en staplad generalisering är det bäst att använda basinlärningsmetodens klassannolikheter som metarepresentation, medan när man använder en icke-linjär metainlärningsmetod för en staplade generaliseringen är det bäst att använda en kombination av basinlärningsmetodens klassannolikheter och den ursprungliga representationen som metarepresentation. Av alla utvärderade sätt att stapla är staplandet med en icke-linjär metainlärningsmetod med en kombination av basinlärningsmetodens klassannolikheter och den ursprungliga representationen som metarepresentation den ansats som presterar bäst i våra experiment över de tio datamängderna.