Le défi du BigData entraîne un besoin pour les algorithmes d'apprentissage automatisé de s'adapter aux données de grande dimension et de devenir plus efficace. Récemment, une nouvelle direction de recherche est apparue qui consiste à analyser les méthodes d’apprentissage dans le régime moderne où le nombre n et la dimension p des données sont grands et du même ordre. Par rapport au régime conventionnel où n>>p, le régime avec n,p sont grands et comparables est particulièrement intéressant, car les performances d’apprentissage dans ce régime restent sensibles à l’ajustement des hyperparamètres, ouvrant ainsi une voie à la compréhension et à l’amélioration des techniques d’apprentissage pour ces données de grande dimension.L'approche technique de cette thèse s'appuie sur des outils avancés de statistiques de grande dimension, nous permettant de mener des analyses allant au-delà de l'état de l’art. La première partie de la thèse est consacrée à l'étude de l'apprentissage semi-supervisé sur des grandes données. Motivés par nos résultats théoriques, nous proposons une alternative supérieure à la méthode semi-supervisée de régularisation laplacienne. Les méthodes avec solutions implicites, comme les SVMs et la régression logistique, sont ensuite étudiées sous des modèles de mélanges réalistes, fournissant des détails exhaustifs sur le mécanisme d'apprentissage. Plusieurs conséquences importantes sont ainsi révélées, dont certaines sont même en contradiction avec la croyance commune. / The BigData challenge induces a need for machine learning algorithms to evolve towards large dimensional and more efficient learning engines. Recently, a new direction of research has emerged that consists in analyzing learning methods in the modern regime where the number n and the dimension p of data samples are commensurately large. Compared to the conventional regime where n>>p, the regime with large and comparable n,p is particularly interesting as the learning performance in this regime remains sensitive to the tuning of hyperparameters, thus opening a path into the understanding and improvement of learning techniques for large dimensional datasets.The technical approach employed in this thesis draws on several advanced tools of high dimensional statistics, allowing us to conduct more elaborate analyses beyond the state of the art. The first part of this dissertation is devoted to the study of semi-supervised learning on high dimensional data. Motivated by our theoretical findings, we propose a superior alternative to the standard semi-supervised method of Laplacian regularization. The methods involving implicit optimizations, such as SVMs and logistic regression, are next investigated under realistic mixture models, providing exhaustive details on the learning mechanism. Several important consequences are thus revealed, some of which are even in contradiction with common belief.

Les Interfaces Cerveau-Machine (ICM) ont été appliquées avec succès aussi bien dans le domaine clinique que pour l'amélioration de la vie quotidienne de patients avec des handicaps. En tant que composante essentielle, le module de traitement du signal détermine nettement la performance d'un système ICM. Nous nous consacrons à améliorer les stratégies de traitement du signal du point de vue de l'apprentissage de la machine. Tout d'abord, nous avons développé un algorithme basé sur les SVM transductifs couplés aux noyaux multiples afin d'intégrer différentes vues des données (vue statistique ou vue géométrique) dans le processus d'apprentissage. Deuxièmement, nous avons proposé une version enligne de l'apprentissage multi-noyaux dans le cas supervisé. Les résultats expérimentaux montrent de meilleures performances par rapport aux approches classiques. De plus, l'algorithme proposé permet de sélectionner automatiquement les canaux de signaux EEG utiles grâce à l'apprentissage multi-noyaux.Dans la dernière partie, nous nous sommes attaqués à l'amélioration du module de traitement du signal au-delà des algorithmes d'apprentissage automatique eux-mêmes. En analysant les données ICM hors-ligne, nous avons d'abord confirmé qu'un modèle de classification simple peut également obtenir des performances satisfaisantes en effectuant une sélection de caractéristiques (et/ou de canaux). Nous avons ensuite conçu un système émotionnel ICM par en tenant compte de l'état émotionnel de l'utilisateur. Sur la base des données de l'EEG obtenus avec différents états émotionnels, c'est-à -dire, positives, négatives et neutres émotions, nous avons finalement prouvé que l'émotion affectait les performances ICM en utilisant des tests statistiques. Cette partie de la thèse propose des bases pour réaliser des ICM plus adaptées aux utilisateurs. / Brain-computer Interface (BCI) has achieved numerous successful applications in both clinicaldomain and daily life amelioration. As an essential component, signal processing determines markedly the performance of a BCI system. In this thesis, we dedicate to improve the signal processing strategy from perspective of machine learning strategy. Firstly, we proposed TSVM-MKL to explore the inputs from multiple views, namely, from statistical view and geometrical view; Secondly, we proposed an online MKL to reduce the computational burden involved in most MKL algorithm. The proposed algorithms achieve a better classifcation performance compared with the classical signal kernel machines, and realize an automatical channel selection due to the advantages of MKL algorithm. In the last part, we attempt to improve the signal processing beyond the machine learning algorithms themselves. We first confirmed that simple classifier model can also achieve satisfying performance by careful feature (and/or channel) selection in off-line BCI data analysis. We then implement another approach to improve the BCI signal processing by taking account for the user's emotional state during the signal acquisition procedure. Based on the reliable EEG data obtained from different emotional states, namely, positive, negative and neutral emotions, we perform strict evaluation using statistical tests to confirm that the emotion does affect BCI performance. This part of work provides important basis for realizing user-friendly BCIs.

Sequence labeling is a type of machine learning problem that involves as- signing a label to each sequence member. Deep learning has shown good per- formance for this problem. However, one disadvantage of this approach is its requirement of having a large amount of labeled data. Semi-supervised learning mitigates this problem by using cheaper unlabeled data together with labeled data. Currently, usage of semi-supervised deep learning for sequence labeling is limited. Therefore, the focus of this thesis is on the application of semi-super- vised deep learning in sequence labeling. Existing semi-supervised deep learning approaches are examined, and approaches for sequence labeling are proposed. The proposed approaches were implemented and experimentally evaluated on named-entity recognition and part-of-speech tagging tasks.

Electrical scooters have gained a lot of attention and popularity among commuters all around the world since they entered the market. After all, electrical scooters have shown to be efficient and cost-effective mode of transportation for commuters and travelers. As of today electrical scooters have firmly established themselves in the micromobility industry, with an increasing global demand.  Although, as the industry is booming so are the accidents as well as getting into dangerous situations of riding electrical scooters. There is a growing concern regarding the safety of the scooters where more and more people are getting injured.   This research focuses on training a computer vision model using semi-supervised learning to help detect traffic rule violations and also prevent collisions for people using electrical scooters. However, applying a computer vision model on an embedded system can be challenging due to the limited capabilities of the hardware. This is where the model can enable post-training quantizations. This thesis examines which post-training quantization has the best performance and if it can perform better compared to the non-quantized model. There are three post-training quantizations that are being applied to the model, dynamic range, full integer and float16 post-training quantizations. The results showed that the non-quantized model achieved a mean average precision (mAP) of 0.03894 with a mean average training and validation loss of 22.10 and 28.11. The non-quantized model was compared with the three post-training quantizations in terms of mAP where the dynamic range post-training quantization achieve the best performance with a mAP of 0.03933.

In this research the focus is on data-centric AI with a specific concentration on data quality evaluation and improvement for machine learning. We first present a practical framework for data quality evaluation and improvement, using a legal domain as a case study and build a corpus for legal argument mining. We first created an initial corpus with 4,937 instances that were manually labeled. We define five data quality evaluation dimensions: comprehensiveness, correctness, variety, class imbalance, and duplication, and conducted a quantitative evaluation on these dimensions for the legal dataset and two existing datasets in the medical domain for medical concept normalization. The first group of experiments showed that class imbalance and insufficient training data are the two major data quality issues that negatively impacted the quality of the system that was built on the legal corpus. The second group of experiments showed that the overlap between the test datasets and the training datasets, which we defined as "duplication," is the major data quality issue for the two medical corpora. We explore several widely used machine learning methods for data quality improvement. Compared to pseudo-labeling, co-training, and expectation-maximization (EM), generative adversarial network (GAN) is more effective for automated data augmentation, especially when a small portion of labeled data and a large amount of unlabeled data is available. The data validation process, the performance improvement strategy, and the machine learning framework for data evaluation and improvement discussed in this dissertation can be used by machine learning researchers and practitioners to build high-performance machine learning systems. All the materials including the data, code, and results will be released at: https://github.com/haihua0913/dissertation-dqei.

<p>  </p> <p>Large-scale deep learning models have reached previously unattainable performance for various tasks. However, the ever-growing resource consumption of neural networks generates large carbon footprint, brings difficulty for academics to engage in research and stops emerging economies from enjoying growing Artificial Intelligence (AI) benefits. To further scale AI to bring more benefits, two major challenges need to be solved. Firstly, even though large-scale deep learning models achieved remarkable success, their performance is still not satisfactory when fine-tuning with only a handful of examples, thereby hindering widespread adoption in real-world applications where a large scale of labeled data is difficult to obtain. Secondly, current machine learning models are still mainly designed for tasks in closed environments where testing datasets are highly similar to training datasets. When the deployed datasets have distribution shift relative to collected training data, we generally observe degraded performance of developed models. How to build adaptable models becomes another critical challenge. To address those challenges, in this dissertation, we focus on two topics: few-shot learning and domain adaptation, where few-shot learning aims to learn tasks with limited labeled data and domain adaption address the discrepancy between training data and testing data. In Part 1, we show our few-shot learning studies. The proposed few-shot solutions are built upon large-scale language models with evolutionary explorations from improving supervision signals, incorporating unlabeled data and improving few-shot learning abilities with lightweight fine-tuning design to reduce deployment costs. In Part 2, domain adaptation studies are introduced. We develop a progressive series of domain adaption approaches to transfer knowledge across domains efficiently to handle distribution shifts, including capturing common patterns across domains, adaptation with weak supervision and adaption to thousands of domains with limited labeled data and unlabeled data. </p>

<p>Graph diffusion is the process of spreading information from one or few nodes to the rest of the graph through edges. The resulting distribution of the information often implies latent structure of the graph where nodes more densely connected can receive more signal. This makes graph diffusions a powerful tool for local clustering, which is the problem of finding a cluster or community of nodes around a given set of seeds. Most existing literatures on using graph diffusions for local graph clustering are linear diffusions as their dynamics can be fully interpreted through linear systems. They are also referred as eigenvector, spectral, or random walk based methods. While efficient, they often have difficulty capturing the correct boundary of a target label or target cluster. On the contrast, maxflow-mincut based methods that can be thought as 1-norm nonlinear variants of the linear diffusions seek to "improve'' or "refine'' a given cluster and can often capture the boundary correctly. However, there is a lack of literature to adopt them for problems such as community detection, local graph clustering, semi-supervised learning, etc. due to the complexity of their formulation. We addressed these issues by performing extensive numerical experiments to demonstrate the performance of flow-based methods in graphs from various sources. We also developed an efficient LocalGraphClustering Python Package that allows others to easily use these methods in their own problems. While studying these flow-based methods, we find that they cannot grow from small seed set. Although there are hybrid procedures that incorporate ideas from both linear diffusions and flow-based methods, they have many hard to set parameters. To tackle these issues, we propose a simple generalization of the objective function behind linear diffusion and flow-based methods which we call generalized local graph min-cut problem. We further show that by involving p-norm in this cut problem, we can develop a nonlinear diffusion procedure that can find local clusters from small seed set and capture the correct boundary simultaneously. Our method can be thought as a nonlinear generalization of the Anderson-Chung-Lang push procedure to approximate a personalized PageRank vector efficiently and is a strongly local algorithm-one whose runtime depends on the size of the output rather than the size of the graph. We also show that the p-norm cut functions improve on the standard Cheeger inequalities for linear diffusion methods. We further extend our generalized local graph min-cut problem and the corresponding diffusion solver to hypergraph-based machine learning problems. Although many methods for local graph clustering exist, there are relatively few for localized clustering in hypergraphs. Moreover, those that exist often lack flexibility to model a general class of hypergraph cut functions or cannot scale to large problems. Our new hypergraph diffusion method on the other hand enables us to compute with a wide variety of cardinality-based hypergraph cut functions and still maintains the strongly local property. We also show that the clusters found by solving the new objective function satisfy a Cheeger-like quality guarantee.</p> <p>Besides clustering, recent work on graph-based learning often focuses on node embeddings and graph neural networks. Although these GNN based methods can beat traditional ones especially when node attributes data is available, it is challenging to understand them because they are highly over-parameterized. To solve this issue, we propose a novel framework that combines topological data analysis and diffusion to transform the complex prediction space into human understandable pictures. The method can be applied to other datasets not in graph formats and scales up to large datasets across different domains and enable us to find many useful insights about the data and the model.</p>

In many relevant scenarios, there is an imbalance between abundant unlabeled data and scarce labeled data to train predictive models. Semi-Supervised Learning and Active Learning are two distinct approaches to deal with this issue. The first one directly uses the unlabeled data to improve model parameter learning, while the second performs a smart choice of unlabeled points to be sent to an annotator, or oracle, which can label these points and increase the labeled training set. In this context, Generative Models are highly appropriate, since they internally represent the data generating process, naturally benefiting from data samples independently of the presence of labels. This Thesis proposes Expectation-Maximization with Density-Weighted Entropy, a novel active semi-supervised learning framework tailored towards generative models. The method is theoretically explored and experiments are conducted to evaluate its application to Gaussian Mixture Models and Multinomial Mixture Models. Based on its partial success, several questions are raised and discussed as to identify possible improvements and decide which shortcomings need to be dealt with before the method is considered robust and generally applicable. / I många relevanta scenarier finns det en obalans mellan god tillgång på oannoterad data och sämre tillgång på annoterad data för att träna prediktiva modeller. Semi-Övervakad Inlärning och Aktiv Inlärning är två distinkta metoder för att hantera denna fråga. Den första använder direkt oannoterad data för att förbättra inlärningen av modellparametrar, medan den andra utför ett smart val av oannoterade punkter som ska skickas till en annoterare eller ett orakel, som kan annotera dessa punkter och öka det annoterade träningssetet. I detta sammanhang är Generativa Modeller mycket lämpliga eftersom de internt representerar data-genereringsprocessen och naturligt gynnas av dataexempel oberoende av närvaron av etiketter. Denna Masteruppsats föreslår Expectation-Maximization med Density-Weighted Entropy, en ny aktiv semi-övervakad inlärningsmetod som är skräddarsydd för generativa modeller. Metoden utforskas teoretiskt och experiment genomförs för att utvärdera dess tillämpning på Gaussiska Mixturmodeller och Multinomiala Mixturmodeller. Baserat på dess partiella framgång ställs och diskuteras flera frågor för att identifiera möjliga förbättringar och avgöra vilka brister som måste hanteras innan metoden anses robust och allmänt tillämplig.

In this thesis, the use of unsupervised and semi-supervised machine learning techniques was analyzed as potential tools for anomaly detection in the sensor network that the electrical system in a Scania truck is comprised of. The experimentation was designed to analyse the need for both point and contextual anomaly detection in this setting. For the point anomaly detection the method of Isolation Forest was experimented with and for contextual anomaly detection two different recurrent neural network architectures using Long Short Term Memory units was relied on. One model was simply a many to one regression model trained to predict a certain signal, while the other was an encoder-decoder network trained to reconstruct a sequence. Both models were trained in an semi-supervised manner, i.e. on data that only depicts normal behaviour, which theoretically should lead to a performance drop on abnormal sequences resulting in higher error terms. In both setting the parameters of a Gaussian distribution were estimated using these error terms which allowed for a convenient way of defining a threshold which would decide if the observation would be flagged as anomalous or not. Additional experimentation's using an exponential weighted moving average over a number of past observations to filter the signal was also conducted. The models performance on this particular task was very different but the regression model showed a lot of promise especially when combined with a filtering preprocessing step to reduce the noise in the data. However the model selection will always be governed by the nature the particular task at hand so the other methods might perform better in other settings. / I den här avhandlingen var användningen av oövervakad och halv-övervakad maskininlärning analyserad som ett möjligt verktyg för att upptäcka avvikelser av anomali i det sensornätverk som elektriska systemet en Scanialastbil består av. Experimentet var konstruerat för att analysera behovet av både punkt och kontextuella avvikelser av anomali i denna miljö. För punktavvikelse av anomali var metoden Isolation Forest experimenterad med och för kontextuella avvikelser av anomali användes två arkitekturer av återkommande neurala nätverk. En av modellerna var helt enkelt många-till-en regressionmodell tränad för att förutspå ett visst märke, medan den andre var ett kodare-avkodare nätverk tränat för att rekonstruera en sekvens.Båda modellerna blev tränade på ett halv-övervakat sätt, d.v.s. på data som endast visar normalt beteende, som teoretiskt skulle leda till minskad prestanda på onormala sekvenser som ger ökat antal feltermer. I båda fallen blev parametrarna av en Gaussisk distribution estimerade på grund av dessa feltermer som tillåter ett bekvämt sätt att definera en tröskel som skulle bestämma om iakttagelsen skulle bli flaggad som en anomali eller inte. Ytterligare experiment var genomförda med exponentiellt viktad glidande medelvärde över ett visst antal av tidigare iakttagelser för att filtera märket. Modellernas prestanda på denna uppgift var välidt olika men regressionmodellen lovade mycket, särskilt kombinerad med ett filterat förbehandlingssteg för att minska bruset it datan. Ändå kommer modelldelen alltid styras av uppgiftens natur så att andra metoder skulle kunna ge bättre prestanda i andra miljöer.

Knowledge Graphs are fascinating concepts in machine learning as they can hold usefully structured information in the form of entities and their relations. Despite the valuable applications of such graphs, most knowledge bases remain incomplete. This missing information harms downstream applications such as information retrieval and opens a window for research in statistical relational learning tasks such as node classification and link prediction. This work proposes a deep learning framework based on existing relational convolutional (R-GCN) layers to learn on highly multi-relational data characteristic of realistic knowledge graphs for node property classification tasks. We propose a deep and improved variant, Deep-RGCNs, with dense and residual skip connections between layers. These skip connections are known to be very successful with popular deep CNN-architectures such as ResNet and DenseNet. In our experiments, we investigate and compare the performance of Deep-RGCN with different baselines on multi-relational graph benchmark datasets, AIFB and MUTAG, and show how the deep architecture boosts the performance in the task of node property classification. We also study the training performance of Deep-RGCNs (with N layers) and discuss the gradient vanishing and over-smoothing problems common to deeper GCN architectures.