<div><div><div><p>Face redaction is needed by law enforcement and mass media outlets to guarantee privacy. In this thesis, a performance analysis of several face redaction/obscuration methods, such as blurring and pixelation is presented. The analysis is based on various threat models and obscuration attackers to achieve a comprehensive evaluation. We show that the traditional blurring and pixelation methods cannot guarantee privacy. To provide a more secured privacy protection, we propose two novel obscuration methods that are based on the generative adversarial networks. The proposed methods not only remove the identifiable information, but also preserve the non-identifiable facial information (as known as the utility information), such as expression, age, skin tone and gender.</p><p>We also propose methods for change detection in satellite imagery. In this thesis, we consider two types of building changes: 2D appearance change and 3D height change. We first present a model with an attention mechanism to detect the building appearance changes that are caused by natural disasters. Furthermore, to detect the changes of building height, we present a height estimation model that is based on building shadows and solar angles without relying on height annotation. Both change detection methods require good building segmentation performance, which might be hard to achieve for the low-quality images, such as off-nadir images. To solve this issue, we use uncertainty modeling and satellite imagery metadata to achieve accurate building segmentation for the noisy images that are taken from large off-nadir angles.</p></div></div></div>

Generation of realistic skydome images We aim to generate realistic images of the sky with clouds using generative adversarial networks (GANs). We explore two GAN architectures, ProGAN and StyleGAN, and find that StyleGAN produces significantly better results. We also propose a novel architecture SuperGAN which aims to generate images at very high resolutions, which cannot be efficiently handled using state-of-art architectures. 1

We expand the scope of image-to-image translation to include more distinct image domains, where the image sets have analogous structures, but may not share object types between them. Semantic-Driven Unsupervised Image-to-Image Translation for Distinct Image Domains (SUNIT) is built to more successfully translate images in this setting, where content from one domain is not found in the other. Our method trains an image translation model by learning encodings for semantic segmentations of images. These segmentations are translated between image domains to learn meaningful mappings between the structures in the two domains. The translated segmentations are then used as the basis for image generation. Beginning image generation with encoded segmentation information helps maintain the original structure of the image. We qualitatively and quantitatively show that SUNIT improves image translation outcomes, especially for image translation tasks where the image domains are very distinct.

La prédominance d’algorithmes de prise de décision, qui sont souvent basés sur desmodèles issus de l’apprentissage machine, soulève des enjeux importants en termes de ladiscrimination et du manque d’équité par ceux-ci ainsi que leur impact sur le traitement degroupes minoritaires ou sous-représentés. Cela a toutefois conduit au développement de tech-niques dont l’objectif est de mitiger ces problèmes ainsi que les les difficultés qui y sont reliées. Dans ce mémoire, nous analysons certaines de ces méthodes d’amélioration de l’équitéde type «pré-traitement» parmi les plus récentes, et mesurons leur impact sur le compromiséquité-utilité des données transformées. Plus précisément, notre focus se fera sur troistechniques qui ont pour objectif de cacher un attribut sensible dans un ensemble de données,dont deux basées sur les modèles générateurs adversériaux (LAFTR [67] et GANSan [6])et une basée sur une transformation déterministe et les fonctions de densités (DisparateImpact Remover [33]). Nous allons premièrement vérifier le niveau de contrôle que cestechniques nous offrent quant au compromis équité-utilité des données. Par la suite, nousallons investiguer s’il est possible d’inverser la transformation faite aux données par chacunde ces algorithmes en construisant un auto-encodeur sur mesure qui tentera de reconstruireles données originales depuis les données transformées. Finalement, nous verrons qu’unacteur malveillant pourrait, avec les données transformées par ces trois techniques, retrouverl’attribut sensible qui est censé être protégé avec des algorithmes d’apprentissage machinede base. Une des conclusions de notre recherche est que même si ces techniques offrentdes garanties pratiques quant à l’équité des données produites, il reste souvent possible deprédire l’attribut sensible en question par des techniques d’apprentissage, ce qui annulepotentiellement toute protection que la technique voulait accorder, créant ainsi de sérieuxdangers au niveau de la vie privée. / The prevalence of decision-making algorithms, based on increasingly powerful patternrecognition machine learning algorithms, has brought a growing wave of concern about dis-crimination and fairness of those algorithm predictions as well as their impacts on equity andtreatment of minority or under-represented groups. This in turn has fuelled the developmentof new techniques to mitigate those issues and helped outline challenges related to such issues. n this work, we analyse recent advances in fairness enhancing pre-processing techniques,evaluate how they control the fairness-utility trade-off and the dataset’s ability to be usedsuccessfully in downstream tasks. We focus on three techniques that attempt to hide asensitive attribute in a dataset, two based onGenerative Adversarial Networksarchitectures(LAFTR [67] and GANSan [6]), and one deterministic transformation of dataset relyingon density functions (Disparate Impact Remover [33]). First we analyse the control overthe fairness-utility trade-off each of these techniques offer. We then attempt to revertthe transformation on the data each of these techniques applied using a variation of anauto-encoder built specifically for this purpose, which we calledreconstructor. Lastly wesee that even though these techniques offer practical guarantees of specific fairness metrics,basic machine learning classifiers are often able to successfully predict the sensitive attributefrom the transformed data, effectively enabling discrimination. This creates what we believeis a major issue in fairness-enhancing technique research that is in large part due to intricaterelationship between fairness and privacy.

No description available.

The forestry and the sawmill industry have a lot of incoming and outgoing piles of wood. It's important to maintain quality and efficiency. This motivates an examination of whether machine learning- or more specifically, anomaly detection techniques can be implemented and used to detect faulty shipments. This thesis presents and evaluates some computer vision techniques and some deep learning techniques. Deep learning can be divided into groups; supervised, semi-supervised and unsupervised. In this thesis, all three groups were examined and it covers supervised methods such as Convolutional Neural Networks, semi-supervised methods such as a modified Convolutional Autoencoder (CAE) and lastly, an unsupervised technique such as Generative Adversarial Network (GAN) was being tested and evaluated.  A version of a GAN model proved to perform best for this thesis in terms of the accuracy of faulty detecting shipments with an accuracy rate of 68.2% and 79.8\% overall, which was satisfactory given the problems that were discovered during the progress of the thesis.

Motivated by successes in the image generation domain, this thesis presents a novel Hybrid VQ-VAE (H-VQ-VAE) approach for generating realistic synthetic time series data with categorical features. The primary motivation behind this work is to address the limitations of existing generative models in accurately capturing the underlying structure and patterns of time series data, especially when dealing with categorical features.  Our proposed H-VQ-VAE model builds upon the foundation of the VQ-VAE architecture and consists of two separate VQ-VAEs: the whole VQ-VAE and the sliding VQ-VAE. Both models share a ResNet-based architecture with conv1d layers to effectively capture the temporal structure within the time series data. The whole VQ-VAE focuses on entire sequences of data to learn relationships between categorical and numerical features, while the sliding VQ-VAE exclusively processes numerical features using a sliding window approach. We conducted experiments on multiple datasets to evaluate the performance of our H-VQ-VAE model in comparison with the original VQ-VAE and TimeGAN models. Our evaluation used a train-on-real and test-on-synthetic approach, focusing on metrics such as Mean Absolute Error (MAE) and Explained Variance (EV). The H-VQ-VAE model achieved a 25-50% better MAE for numerical features compared to the VQ-VAE and outperformed TimeGAN by 45-75% on the complex dataset indicating its effectiveness in capturing the underlying structure and patterns of the time series data. In conclusion, the H-VQ-VAE model offers a promising approach for generating realistic synthetic time series data with categorical features, with potential applications in various fields where accurate data generation is crucial.

Style transfer using Generative adversarial networks (GANs) has been successful in recent publications. One field in style transfer is music style transfer, in which a piece of music is transformed in some way, be it through genre-, harmonic-, rhythmic transfer, etc. In this thesis, we have performed genre style transfer using a CycleGAN architecture and symbolic representation of data. Previous work using the same architecture and representation has focused solely on transferring the arrangement of the notes (composition). We have improved this work by including the transfer of multiple instruments (timbre) to create more convincing results. Additional discriminators were added to the CycleGAN architecture to achieve this, and they are individually tasked with enforcing the timbre and composition of a song. Previous works have also used variable autoencoders (VAEs) with sequential data representation for style transfer. The use of VAEs for genre style transfer using symbolic data representation instead of sequential was explored, and recommendations for future work include omitting faults found during exploration. Two different classifiers were created to evaluate the results of the CycleGAN model. One uses symbolic representation, in which all instruments are merged into one, thus evaluating the composition of the generated songs. The other classifier uses a spectrogram representation which evaluates the transfer as a whole, both timbre and composition. The evaluation of the improved CycleGAN model using the classifiers showed that it could perform genre style transfer successfully even when adding timbre to the style transfer.

Phase transitions are fundamental phenomena in (quantum) many-body systems. They are associated with changes in the macroscopic physical properties of the system in response to the alteration in the conditions controlled by one or more parameters, like temperature or coupling constants. Quantum phase transitions are particularly intriguing as they reveal new insights into the fundamental nature of matter and the laws of physics. The study of phase transitions in such systems is crucial in aiding our understanding of how materials behave in extreme conditions, which are difficult to replicate in laboratory, and also the behavior of exotic states of matter with unique and potentially useful properties like superconductors and superfluids. Moreover, this understanding has other practical applications and can lead to the development of new materials with specific properties or more efficient technologies, such as quantum computers. Hence, detecting the transition point from one phase of matter to another and constructing the corresponding phase diagram is of great importance for examining many-body systems and predicting their response to external perturbations. Traditionally, phase transitions have been identified either through analytical methods like mean field theory or numerical simulations. The pinpointing of the critical value normally involves the measure of specific quantities such as local observables, correlation functions, energy gaps, etc. reflecting the changes in the physics through the transition. However, the latter approach requires prior knowledge of the system to calculate the order parameter of the transition, which is uniquely associated to its universality class. Recently, another method has gained more and more attention in the physics community. By using raw and very general representative data of the system, one can resort to machine learning techniques to distinguish among patterns within the data belonging to different phases. The relevance of these techniques is rooted in the ability of a properly trained machine to efficiently process complex data for the sake of pursuing classification tasks, pattern recognition, generating brand new data and even developing decision processes. The aim of this thesis is to explore phase transitions from this new and promising data-centric perspective. On the one hand, our work is focused on the developement of new machine learning architectures using state-of-the-art and interpretable models. On the other hand, we are interested in the study of the various possible data which can be fed to the artificial intelligence model for the mapping of a quantum many-body system phase diagram. Our analysis is supported by numerical examples obtained via matrix-product-states (MPS) simulations for several one-dimensional zero-temperature systems on a lattice such as the XXZ model, the Extended Bose-Hubbard model (EBH) and the two-species Bose Hubbard model (BH2S). In Part I, we provide a general introduction to the background concepts for the understanding of the physics and the numerical methods used for the simulations and the analysis with deep learning. In Part II, we first present the models of the quantum many-body systems that we study. Then, we discuss the machine learning protocol to identify phase transitions, namely anomaly detection technique, that involves the training of a model on a dataset of normal behavior and use it to recognize deviations from this behavior on test data. The latter can be applied for our purpose by training in a known phase so that, at test-time, all the other phases of the system are marked as anomalies. Our method is based on Generative Adversarial Networks (GANs) and improves the networks adopted by the previous works in the literature for the anomaly detection scheme taking advantage of the adversarial training procedure. Specifically, we train the GAN on a dataset composed of bipartite entanglement spectra (ES) obtained from Tensor Network simulations for the three aforementioned quantum systems. We focus our study on the detection of the elusive Berezinskii-Kosterlitz-Thouless (BKT) transition that have been object of intense theoretical and experimental studies since its first prediction for the classical two-dimensional XY model. The absence of an explicit symmetry breaking and its gappless-to-gapped nature which characterize such a transition make the latter very subtle to be detected, hence providing a challenging testing ground for the machine-driven method. We train the GAN architecture on the ES data in the gapless side of BKT transition and we show that the GAN is able to automatically distinguish between data from the same phase and beyond the BKT. The protocol that we develop is not supposed to become a substitute to the traditional methods for the phase transitions detection but allows to obtain a qualitative map of a phase diagram with almost no prior knowledge about the nature and the arrangement of the phases -- in this sense we refer to it as agnostic -- in an automatic fashion. Furthermore, it is very general and it can be applied in principle to all kind of representative data of the system coming both from experiments and numerics, as long as they have different patterns (even hidden to the eye) in different phases. Since the kind of data is crucially linked with the success of the detection, together with the ES we investigate another candidate: the probability density function (PDF) of a globally U(1) conserved charge in an extensive sub-portion of the system. The full PDF is one of the possible reductions of the ES which is known to exhibit relations and degeneracies reflecting very peculiar aspects of the physics and the symmetries of the system. Its patterns are often used to tell different kinds of phases apart and embed information about non-local quantum correlations. However, the PDF is measurable, e.g. in quantum gas microscopes experiments, and it is quite general so that it can be considered not only in the cases of the study but also in other systems with different symmetries and dimensionalities. Both the ES and the PDF can be extracted from the simulation of the ground state by dividing the one-dimensional chain into two complementary subportions. For the EBH we calculate the PDF of the bosonic occupation number in a wide range of values of the couplings and we are able to reproduce the very rich phase diagram containing several phases (superfluid, Mott insulator, charge density wave, phase separation of supersolid and superfluid and the topological Haldane insulator) just with an educated gaussian fit of the PDF. Even without resorting to machine learning, this analysis is instrumental to show the importance of the experimentally accessible PDF for the task. Moreover, we highlight some of its properties according to the gapless and gapped nature of the ground state which require a further investigation and extension beyond zero-temperature regimes and one-dimensional systems. The last chapter of the results contains the description of another architecture, namely the Concrete Autoencoder (CAE) which can be used for detecting phase transitions with the anomaly detection scheme while being able to automatically learn what the most relevant components of the input data are. We show that the CAE can recognize the important eigenvalues out of the entire ES for the EBH model in order to characterize the gapless phase. Therefore the latter architecture can be used to provide not only a more compact version of the input data (dimensionality reduction) -- which can improve the training -- but also some meaningful insights in the spirit of machine learning interpretability. In conclusion, in this thesis we describe two advances in the solution to the problem of phase recognition in quantum many-body systems. On one side, we improve the literature standard anomaly detection protocol for an automatic and agnostic identification of the phases by employing the GAN network. Moreover, we implement and test an explainable model which can make the interpretation of the results easier. On the other side we put the focus on the PDF as a new candidate quantity for the scope of discerning phases of matter. We show that it contains a lot of information about the many-body state being very general and experimentally accessible.

The steep growth of video-games is demanding a higher amount of characters in the games. The process of generating characters is very expensive and time consuming. Consequently, this process doesn’t cover the current demands and could be optimized by developing a generative model able to synthesize high- quality 3D avatar faces within minutes. This model would result in drastic gains for gaming companies. Therefore, the aim of this project is to implement a model able to generate realistic 3D avatar faces by generating texture and shape when a limited amount of data is given (&lt;1k samples). This type of model is called 3D Morphable Model and it will also learn the correlation between shape and texture in order to generate consistent results. In order to achieve this final model, which is called joint model, individual models for texture and shape are also developed. The three type of models are built upon StyleGAN2-ADA architecture. The final design of the joint model has three discriminators: a joint discriminator to ensure consistency and two individual discriminators to have good quality for shape and texture. This model was inspired from [1]. The experiments show that the best texture model uses the augmentation techniques introduced in StyleGAN2-ADA. The experiments over the joint model prove that having just one discriminator is not enough to generate good quality results. On the other hand, the joint model with three discriminators give good quality and coherent results. In addition, this joint model outperforms the results of the shape model when training the model with the same number of samples, 969 samples. This model shows a promising path for further improvements. / Den kraftiga ökningen av videospel kräver ett större antal karaktärer i spelen. Processen att skapa karaktärer är mycket dyr och tidskrävande. Denna process täcker därför inte den nuvarande efterfrågan och skulle kunna optimeras genom att utveckla en generativ modell som kan syntetisera högkvalitativa 3D-avataransikten av hög kvalitet på några minuter. Denna modell skulle leda till drastiska vinster för spelföretagen. Syftet med detta projekt är därför att implementera en modell som kan generera realistiska 3D-avataransikten genom att generera textur och form när en begränsad mängd data ges (&lt;1k samplingar). Denna typ av modell kallas 3D Morphable Model och den kommer också att lära sig korrelationen mellan form och textur för att generera konsekventa resultat. För att uppnå denna slutliga modell, som kallas gemensam modell, utvecklas också individuella modeller för textur och form. De tre typerna av modeller bygger på StyleGAN2- ADA-arkitekturen. Den slutliga utformningen av den gemensamma modellen har tre diskriminatorer: en gemensam diskriminator för att säkerställa konsistens och två individuella diskriminatorer för att uppnå god kvalitet för form och textur. Denna modell har inspirerats av [1]. Experimenten visar att den bästa texturmodellen använder de förstärkningstekniker som infördes i StyleGAN2-ADA. Experimenten med den gemensamma modellen visar att det inte räcker med bara en diskriminator för att generera resultat av god kvalitet. Å andra sidan ger den gemensamma modellen med tre diskriminatorer bra kvalitet och sammanhängande resultat. Dessutom överträffar denna gemensamma modell resultaten från formmodellen när modellen tränas med samma antal prov, 969 prov. Denna modell visar en lovande väg för ytterligare förbättringar.