abstract: Endowing machines with the ability to understand digital images is a critical task for a host of high-impact applications, including pathology detection in radiographic imaging, autonomous vehicles, and assistive technology for the visually impaired. Computer vision systems rely on large corpora of annotated data in order to train task-specific visual recognition models. Despite significant advances made over the past decade, the fact remains collecting and annotating the data needed to successfully train a model is a prohibitively expensive endeavor. Moreover, these models are prone to rapid performance degradation when applied to data sampled from a different domain. Recent works in the development of deep adaptation networks seek to overcome these challenges by facilitating transfer learning between source and target domains. In parallel, the unification of dominant semi-supervised learning techniques has illustrated unprecedented potential for utilizing unlabeled data to train classification models in defiance of discouragingly meager sets of annotated data. In this thesis, a novel domain adaptation algorithm -- Domain Adaptive Fusion (DAF) -- is proposed, which encourages a domain-invariant linear relationship between the pixel-space of different domains and the prediction-space while being trained under a domain adversarial signal. The thoughtful combination of key components in unsupervised domain adaptation and semi-supervised learning enable DAF to effectively bridge the gap between source and target domains. Experiments performed on computer vision benchmark datasets for domain adaptation endorse the efficacy of this hybrid approach, outperforming all of the baseline architectures on most of the transfer tasks. / Dissertation/Thesis / Masters Thesis Computer Science 2019

Dealing with large amounts of unlabeled data is a very challenging task. Recently, many different approaches have been proposed to leverage this data for training many machine learning models. Among them, self-supervised learning appears as an efficient solution capable of training powerful and generalizable models. More specifically, instead of relying on human-generated labels, it proposes training objectives that use ``labels'' generated from the data itself, either via data augmentation or by masking the data in some way and trying to reconstruct it. Apart from being able to train models from scratch, self-supervised methods can also be used in specific applications to further improve a pre-trained model. In this thesis, we propose to leverage self-supervised methods in novel ways to tackle different application scenarios. We present four published papers: an open-source library for self-supervised learning that is flexible, scalable, and easy to use; two papers tackling unsupervised domain adaptation in action recognition; and one paper on self-supervised learning for continual learning. The published papers highlight that self-supervised techniques can be leveraged for many scenarios, yielding state-of-the-art results.

In this thesis, we propose a new framework for the generation of training data for machine learning techniques used for classification in communications applications. Machine learning-based signal classifiers do not generalize well when training data does not describe the underlying probability distribution of real signals. The simplest way to accomplish statistical similarity between training and testing data is to synthesize training data passed through a permutation of plausible forms of noise. To accomplish this, a framework is proposed that implements arbitrary channel conditions and baseband signals. A dataset generated using the framework is considered, and is shown to be appropriately sized by having $11\%$ lower entropy than state-of-the-art datasets. Furthermore, unsupervised domain adaptation can allow for powerful generalized training via deep feature transforms on unlabeled evaluation-time signals. A novel Deep Reconstruction-Classification Network (DRCN) application is introduced, which attempts to maintain near-peak signal classification accuracy despite dataset bias, or perturbations on testing data unforeseen in training. Together, feature transforms and diverse training data generated from the proposed framework, teaching a range of plausible noise, can train a deep neural net to classify signals well in many real-world scenarios despite unforeseen perturbations.

<p dir="ltr">Recent advances in Artificial Intelligence (AI) have been markedly accelerated by the convergence of advances in Machine Learning (ML) and the exponential growth in computational power. Within this dynamic landscape, the concept of Domain Adaptation (DA) is dedicated to the seamless transference of knowledge across domains characterized by disparate data distributions. This thesis ventures into the challenging and nuanced terrain of Online Unsupervised Domain Adaptation (OUDA), where the unlabeled data stream arrives from the target domain incrementally and gradually diverges from the source domain. This thesis presents two innovative and complementary approaches -- a manifold-based approach and a time-domain-based approach -- to effectively tackle the intricate OUDA challenges.</p><p dir="ltr">The manifold-based approach seeks to address this gap by incorporating the domain alignment process in an incremental computation manner, and this novel technique leverages the computation of transformation matrices, based on the projection of both source and target data onto the Grassmann manifold. This projection aligns both domains by incrementally minimizing their dissimilarities, effectively ameliorating the divergence between the source and target data. This manifold-based approach capitalizes on the cumulative temporal information within the data stream, utilizing the Incremental Computation of Mean-Subspace (ICMS) technique. This technique efficiently computes the average subspace of target subspaces on the Grassmann manifold, adeptly capturing the evolving dynamics of the data distribution. The alignment process is further fortified by integrating the flow of target subspaces on the manifold. As the target data stream unfolds over time, this approach incorporates this information, yielding robust and adaptive transformation matrices. In addition, the efficient computation of the mean-subspace, closely aligned with the Karcher mean, attests to the computational feasibility of the manifold-based approach, thus, enabling real-time feedback computations for the OUDA problem.</p><p dir="ltr">The time-domain-based approach utilizes the cluster-wise information and its flow information from each time-step to accurately predict target labels in the incoming target data, propagate consistently the class labels to future incoming target data, and efficiently utilize the predicted labels in the target data together with the source data to incrementally update the learning model in a supervised-learning scenario. This process effectively transforms the OUDA problem into a supervised-learning scenario. We leverage a neural-network-based model to align target features, cluster them class-wise and extend them linearly from the origin of the latent space as the time-step progresses. This alignment process enables accurate predictions and target label propagation based on the trajectories of the target features. We achieve target label propagation through the novel Flow-based Hierarchical Optimal Transport (FHOT) method, which considers element-wise, cluster-wise, and distribution-wise correspondences of adjacent target features. The learning model is continuously updated with incoming target data and their predicted labels.</p><p dir="ltr">To comprehensively assess the impact and contributions of these two approaches to the OUDA problem, we conducted extensive experiments across diverse datasets. Our analysis covered each stage of the manifold-based approach, comparing its performance with prior methods in terms of classification accuracy and computational efficiency. The time-domain-based approach was validated through linear feature alignment in the latent space, resulting in accurate label predictions. Notably, the flow-based hierarchical optimal transport technique substantially enhanced classification accuracy, particularly with increasing time-steps. Furthermore, learning model updates using target data and predicted labels significantly improved classification accuracy.</p>

With the rapid growth of computing power, approaches based on deep learning algorithms have achieved remarkable results in solving computer vision classification problems. These performance improvements are achieved by assuming the source and target data are collected from the same probability distribution. However, this assumption is usually too strict to be satisfied in many real-world applications, such as big data analysis, natural language processing, and computer vision classification problems. Because of distribution discrepancies between these domains, directly training the model on the source domain cannot be expected to generate satisfactory results on the target domain. Therefore, the problem of minimizing these data distribution discrepancies is the main challenge with which modern machine learning is now faced. To address this problem, domain adaptation (DA) aims to identify domain-invariant features between two different but related domains. This thesis proposes a state-of-the-art DA approach that overcomes the limitations of traditional DA methods. To capture fine-grained information for each category, I deploy centroid-to-centroid alignment to perform domain adaptation. An Exponential Moving Average strategy (EMA) is used to ensure we can form robust source and target centroids. A Gaussian-uniform mixture model is trained using an Expectation-Maximization (EM) algorithm to infer the robustness of the target pseudo-labels. With the help of target pseudo-labels, I propose two novel types of classifiers: (1) a target-oriented classifier (TO); and (2) a centroid-oriented classifier (CO). Extensive experiments show that these two classifiers exhibit superior performance on a variety of DA benchmarks when compared to standard baseline methods. / Master of Science / Approaches based on deep learning algorithms have achieved remarkable results in solving computer vision classification problems. These performance improvements are achieved by assuming the source and target data are collected from the same probability distribution; however, in many real-world applications, such as big data analysis, natural language processing, and computer vision classification problems, this assumption is usually too strict to be satisfied. For example, these two domains may have the same types of classes, but the objects in each category of these different domains can vary in shape, color, background, or even illumination. Because the probability distributions are slightly mismatched, directly training the model on one domain cannot achieve a satisfactory result on the other domain. To address this problem, domain adaptation (DA) aims to extract common features on both domains to transfer knowledge from one domain to another. In this thesis, I propose a state-of-the-art DA approach that overcomes the limitation of the traditional DA methods. To capture the low-level information of each category, I deploy centroid-to-centroid alignment to perform domain adaptation. An Exponential Moving Average (EMA) strategy is used to ensure the generation of robust centroids. A Gaussian-Uniform Mixture model is trained by using the Expectation-Maximization (EM) algorithm to infer the robustness of the target sample pseudo-labels. With the help of robust target pseudo-labels, I propose two novel types of classifiers: (1) a target-oriented classifier (TO); and (2) a centroid-oriented classifier (CO). Extensive experiments show that the proposed method outperforms traditional baseline methods on various DA benchmarks.

Framgången inom maskininlärning och djupinlärning beror till stor del på stora, annoterade dataset. Att tilldela etiketter till data är väldigt resurskrävande och kan till viss del undvikas genom att utnyttja datans statistiska egenskaper. En maskininlärningsmodell kan lära sig att klassificera bilder från en domän utifrån träningsexempel som innehåller bilder, samt etiketter som berättar vad bilder föreställer. Men vad gör man om datan inte har tilldelade etiketter? En maskininlärningsmodell som lär sig en uppgift utifrån annoterad data från en källdomän, kan med hjälp av information från måldomänen (som inte har tilldelade etiketter), anpassas till att prestera bättre på data från måldomänen. Forskningsområdet som studerar hur man anpassar och generaliserar en modell mellan två olika domäner heter domänanpassning, eller domain adaptation, på engelska.   Detta examensarbete är utfört på Scanias forskningsavdelning för autonom transport och handlar om hur modeller för bildklassificering som tränas på kamerabilder med etiketter, kan anpassas till att få ökad noggrannhet på ett dataset med LiDAR bilder, som inte har etiketter. Två metoder för domänanpassning har jämförts med varandra, samt en model tränad på kameradata genom övervakad inlärning utan domänanpassning. Alla metoder opererar på något vis med ett djupt faltningsnätverk (CNN) där uppgiften är att klassificera bilder utav bilar eller fotgängare. Kovariansen utav datan från käll- och måldomänen är det centrala måttet för domänanpassningsmetoderna i detta projekt. Den första metoden är en så kallad ytlig metod, där själva anpassningsmetoden inte ingår inuti den djupa arkitekturen av modellen, utan är ett mellansteg i processen. Den andra metoden förenar domänanpassningsmetoden med klassificeringen i den djupa arkitekturen. Den tredje modellen består endast utav faltningsnätverket, utan en metod för domänanpassning och används som referens.    Modellen som tränades på kamerabilderna utan en domänanpassningsmetod klassificerar LiDAR-bilderna med en noggrannhet på 63.80%, samtidigt som den ”ytliga” metoden når en noggrannhet på 74.67% och den djupa metoden presterar bäst med 80.73%. Resultaten visar att det är möjligt att anpassa en modell som tränas på data från källdomänen, till att få ökad klassificeringsnoggrannhet i måldomänen genom att använda kovariansen utav datan från de två domänerna. Den djupa metoden för domänanpassning tillåter även användandet utav andra statistiska mått som kan vara mer framgångsrika i att generalisera modellen, beroende på hur datan är fördelad. Överlägsenheten hos den djupa metoden antyder att domänanpassning med fördel kan bäddas in i den djupa arkitekturen så att modelparametrarna blir uppdaterade för att lära sig en mer robust representation utav måldomänen.

Deep Learning models have seen great application in demanding tasks such as machine translation and autonomous driving. However, building such models has proved challenging, both from a computational perspective and due to the requirement of a plethora of annotated data. Moreover, when challenged on new situations or data distributions (target domain), those models may perform inadequately. Such examples are transitioning from one city to another, different weather situations, or changes in sunlight. Unsupervised Domain adaptation (UDA) exploits unlabelled data (easy access) to adapt models to new conditions or data distributions. Inspired by the fact that environmental changes happen gradually, we focus on Online UDA. Instead of directly adjusting a model to a demanding condition, we constantly perform minor adaptions to every slight change in the data, creating a soft transition from the current domain to the target one. To perform gradual adaptation, we utilized state-of-the-art semantic segmentation approaches on increasing rain intensities (25, 50, 75, 100, and 200mm of rain). We demonstrate that deep learning models can adapt substantially better to hard domains when exploiting intermediate ones. Moreover, we introduce a model switching mechanism that allows adjusting back to the source domain, after adaptation, without dropping performance. / Deep Learning-modeller har sett stor tillämpning i krävande uppgifter som maskinöversättning och autonom körning. Att bygga sådana modeller har dock visat sig vara utmanande, både ur ett beräkningsperspektiv och på grund av kravet på en uppsjö av kommenterade data. Dessutom, när de utmanas i nya situationer eller datadistributioner (måldomän), kan dessa modeller prestera otillräckligt. Sådana exempel är övergång från en stad till en annan, olika vädersituationer eller förändringar i solljus. Unsupervised Domain adaptation (UDA) utnyttjar omärkt data (enkel åtkomst) för att anpassa modeller till nya förhållanden eller datadistributioner. Inspirerade av att miljöförändringar sker gradvis, fokuserar vi på Online UDA. Istället för att direkt anpassa en modell till ett krävande tillstånd, gör vi ständigt mindre anpassningar till varje liten förändring i data, vilket skapar en mjuk övergång från den aktuella domänen till måldomänen. För att utföra gradvis anpassning använde vi toppmoderna semantiska segmenteringsmetoder för att öka regnintensiteten (25, 50, 75, 100 och 200 mm regn). Vi visar att modeller för djupinlärning kan anpassa sig betydligt bättre till hårda domäner när man utnyttjar mellanliggande. Dessutom introducerar vi en modellväxlingsmekanism som tillåter justering tillbaka till källdomänen, efter anpassning, utan att tappa prestanda.

[pt] A segmentação de falhas sísmicas apresenta uma tarefa desafiadora edemorada na geofísica, especialmente na exploração e extração de petróleo egás natural. Métodos de Aprendizado Profundo (Deep Learning) têm mostradoum grande potencial para enfrentar esses desafios e oferecem vantagens emcomparação com métodos tradicionais. No entanto, abordagens baseadas emAprendizado Profundo geralmente requerem uma quantidade substancial dedados rotulados, o que contradiz o cenário atual de disponibilidade limitadade dados sísmicos rotulados. Para lidar com essa limitação, pesquisadores têmexplorado a geração de dados sintéticos como uma solução potencial paradados reais não rotulados. Essa abordagem envolve treinar um modelo emdados sintéticos rotulados e, posteriormente, aplicar diretamente ao conjuntode dados real. No entanto, a geração de dados sintéticos encontra o problemade deslocamento de domínio devido à complexidade das situações geológicasdo mundo real, resultando em diferenças na distribuição entre conjuntosde dados sintéticos e reais. Para mitigar o problema de deslocamento dedomínio na detecção de falhas sísmicas, propomos uma nova abordagem queutiliza técnicas de Adaptação de Domínio Não Supervisionada ou UnsupervisedDomain Adaptation (UDA). Nossa proposta envolve o uso de um conjunto dedados sintéticos para treinamento do modelo e sua adaptação a dois conjuntosde dados reais disponíveis publicamente na literatura. As técnicas de UDAescolhidas incluem Maximum Mean Discrepancy (MMD), Domain-AdversarialNeural Networks (DANN) e Fourier Domain Adaptation (FDA). MMD eDANN visam alinhar características em um espaço de características comumde n dimensões, minimizando discrepâncias e aumentando a confusão dedomínio por meio do treinamento adversarial, respectivamente. Por outro lado,FDA transfere o estilo de amostras reais para sintéticas usando TransformadaRápida de Fourier. Para os experimentos, utilizamos uma versão menor doUNet e sua variante Atrous UNet, que incorpora camadas convolucionaisdilatadas em seu gargalo. Além disso, o DexiNed (Dense Extreme InceptionNetwork), um modelo do estado da arte para detecção de bordas, foi empregadopara fornecer uma análise mais abrangente. Além disso, estudamos a aplicaçãode ajuste fino ou fine-tuning em conjuntos de dados rotulados para investigarseu impacto no desempenho, pois muitos estudos o têm utilizado para reduziro deslocamento de domínio.Os resultados finais demonstraram melhorias significativas no desempenho de detecção de falhas ao aplicar técnicas de UDA, com aumento médio deaté 13 por cento em métricas de avaliação como Intersection over Union e F1-score.Além disso, a abordagem proposta obteve detecções mais consistentes de falhassísmicas com menos falsos positivos, indicando seu potencial para aplicações nomundo real. Por outro lado, a aplicação de ajuste fino não demonstrou ganhossignificativos no desempenho, mas reduziu o tempo de treinamento. / [en] Seismic fault segmentation presents a challenging and time-consuming task in geophysics, particularly in the exploration and extraction of oil and natural gas. Deep Learning (DL) methods have shown significant potential to address these challenges and offer advantages compared to traditional methods. However, DL-based approaches typically require a substantial amount of labeled data, which contradicts the current scenario of limited availability of labeled seismic data. To address this limitation, researchers have explored synthetic data generation as a potential solution for unlabeled real data. This approach involves training a model on labeled synthetic data and subsequently applying it directly to the real dataset. However, synthetic data generation encounters the domain shift problem due to the complexity of real-world geological situations, resulting in differences in distribution between synthetic and real datasets. To mitigate the domain shift issue in seismic fault detection, we propose a novel approach utilizing Unsupervised Domain Adaptation (UDA) techniques. Our proposal involves using a synthetic dataset for model training and adapting it to two publicly available real datasets found in the literature. The chosen UDA techniques include Maximum Mean Discrepancy (MMD), Domain-Adversarial Neural Networks (DANN), and Fourier Domain Adaptation (FDA). MMD and DANN aim to align features in a common n-dimensional feature space by minimizing discrepancy and increasing domain confusion through adversarial training, respectively. On the other hand, FDA transfers the style from real to synthetic samples using Fast Fourier Transform. For the experiments, we utilized a smaller version of UNet and its variant Atrous UNet, which incorporates Dilated Convolutional layers in its bottleneck. Furthermore, DexiNed (Dense Extreme Inception Network), a state-of-the-art model for edge detection, was employed to provide a more comprehensive analysis. Additionally, we studied the application of fine-tuning on labeled datasets to investigate its impact on performance, as many studies have employed it to reduce domain shift. The final results demonstrated significant improvements in fault detection performance by applying UDA techniques, with up to a 13 percent increase in evaluation metrics such as Intersection over Union and F1-score on average. Moreover, the proposed approach achieved more consistent detections of seismic faults with fewer false positives, indicating its potential for realworld applications. Conversely, the application of fine-tuning did not show a significant gain in performance but did reduce the training time.