L’expression musicale requiert le contrôle sur quelles notes sont jouées ainsi que comment elles se jouent. Les synthétiseurs audios conventionnels offrent des contrôles expressifs détaillés, cependant au détriment du réalisme. La synthèse neuronale en boîte noire des audios et les échantillonneurs concaténatifs sont capables de produire un son réaliste, pourtant, nous avons peu de mécanismes de contrôle. Dans ce travail, nous introduisons MIDI-DDSP, un modèle hiérarchique des instruments musicaux qui permet tant la synthèse neuronale réaliste des audios que le contrôle sophistiqué de la part des utilisateurs. À partir des paramètres interprétables de synthèse provenant du traitement différentiable des signaux numériques (Differentiable Digital Signal Processing, DDSP), nous inférons les notes musicales et la propriété de haut niveau de leur performance expressive (telles que le timbre, le vibrato, l’intensité et l’articulation). Ceci donne naissance à une hiérarchie de trois niveaux (notes, performance, synthèse) qui laisse aux individus la possibilité d’intervenir à chaque niveau, ou d’utiliser la distribution préalable entraînée (notes étant donné performance, synthèse étant donné performance) pour une assistance créative. À l’aide des expériences quantitatives et des tests d’écoute, nous démontrons que cette hiérarchie permet de reconstruire des audios de haute fidélité, de prédire avec précision les attributs de performance d’une séquence de notes, mais aussi de manipuler indépendamment les attributs étant donné la performance. Comme il s’agit d’un système complet, la hiérarchie peut aussi générer des audios réalistes à partir d’une nouvelle séquence de notes. En utilisant une hiérarchie interprétable avec de multiples niveaux de granularité, MIDI-DDSP ouvre la porte aux outils auxiliaires qui renforce la capacité des individus à travers une grande variété d’expérience musicale. / Musical expression requires control of both what notes are played, and how they are performed. Conventional audio synthesizers provide detailed expressive controls, but at the cost of realism. Black-box neural audio synthesis and concatenative samplers can produce realistic audio, but have few mechanisms for control. In this work, we introduce MIDI-DDSP a hierarchical model of musical instruments that enables both realistic neural audio synthesis and detailed user control. Starting from interpretable Differentiable Digital Signal Processing (DDSP) synthesis parameters, we infer musical notes and high-level properties of their expressive performance (such as timbre, vibrato, dynamics, and articulation). This creates a 3-level hierarchy (notes, performance, synthesis) that affords individuals the option to intervene at each level, or utilize trained priors (performance given notes, synthesis given performance) for creative assistance. Through quantitative experiments and listening tests, we demonstrate that this hierarchy can reconstruct high-fidelity audio, accurately predict performance attributes for a note sequence, independently manipulate the attributes of a given performance, and as a complete system, generate realistic audio from a novel note sequence. By utilizing an interpretable hierarchy, with multiple levels of granularity, MIDI-DDSP opens the door to assistive tools to empower individuals across a diverse range of musical experience.

<p>Deep neural networks have made significant advances in computer vision. However, several challenges limit their real-world applications. For example, domain shifts in vision data degrade model performance; visual appearance variances affect model robustness; it is also non-trivial to extend a model trained on one task to novel tasks; and in many applications, large-scale labeled data are not even available for learning powerful deep models from scratch. This research focuses on improving the transferability of deep features and the efficiency of deep vision model adaptation, leading to enhanced generalization and new capabilities on computer vision tasks. Specifically, we approach these problems from the following two directions: architectural adaptation and label-efficient transferable feature learning. From an architectural perspective, we investigate various schemes that permit network adaptation to be parametrized by multiple copies of sub-structures, distributions of parameter subspaces, or functions that infer parameters from data. We also explore how model adaptation can bring new capabilities, such as continuous and stochastic image modeling, fast transfer to new tasks, and dynamic computation allocation based on sample complexity. From the perspective of feature learning, we show how transferable features emerge from generative modeling with massive unlabeled or weakly labeled data. Such features enable both image generation under complex conditions and downstream applications like image recognition and segmentation. By combining both perspectives, we achieve improved performance on computer vision tasks with limited labeled data, enhanced transferability of deep features, and novel capabilities beyond standard deep learning models.</p>

Magnetic resonance (MR) images together with computed tomography (CT) images are used in many medical practices, such as radiation therapy. To capture those images, patients have to undergo two separate scans: one for the MR image, which involves using strong magnetic fields, and one for the CT image which involves using radiation (x-rays). Another approach is to generate synthetic CT (sCT) images from MR images, thus the patients only have to take one image (the MR image), making the whole process easier and more effcient. One way of generating sCT images is by using generative diffusion models which are a relatively new class in generative models. To this end, this project aims to enquire whether generative diffusion models are capable of generating viable and realistic sCT images from MR images. Firstly, a denoising diffusion probabilistic model (DDPM) with a U-Net backbone neural network is implemented and tested on the MNIST dataset, then it is implemented on a pelvis dataset consisting of 41600 pairs of images, where each pair is made up of an MR image with its respective CT image. The MR images were added at each sampling step in order to condition the sampled sCT images on the MR images. After successful implementation and training, the developed diffusion model got a Fréchet inception distance (FID) score of 14.45, and performed as good as the current state-of-the-art model without any major optimizations to the hyperparameters or to the model itself. The results are very promising and demonstrate the capabilities of this new generative modelling framework.

The usage of image classification in various industries has grown significantly in recentyears. There are however challenges concerning the data used to train such models. Inmany cases the data used in training is often difficult and expensive to obtain. Furthermore,dealing with image data may come with additional problems such as privacy concerns. Inrecent years, synthetic image generation models such as Stable Diffusion has seen signifi-cant improvement. Solely using a textual description, Stable Diffusion is able to generate awide variety of photorealistic images. In addition to textual descriptions, other condition-ing models such as ControlNet has enabled the possibility of additional grounding infor-mation, such as canny edge and segmentation images. This thesis investigates if syntheticimages generated by Stable Diffusion can be used effectively in training an image classifier.To find the most effective method for generating training data, multiple conditioning meth-ods are investigated and evaluated. The results show that it is possible to generate high-quality training data using several conditioning techniques. The best performing methodwas using canny edge grounded images to augment already existing data. Extending twoclasses with additional synthetic data generated by the best performing method, achievedthe highest average F1-score increase of 0.85 percentage points compared with a baselinesolely trained on real images.

This thesis explores the nature of generalization in deep learning and several settings in which it fails. In particular, deep neural networks can struggle to generalize in settings with limited data, insufficient supervision, challenging long-range dependencies, or complex structure and subsystems. This thesis explores the nature of these challenges for generalization in deep learning and presents several algorithms which seek to address these challenges. In the first article, we show how training with interpolated hidden states can improve generalization and calibration in deep learning. We also introduce a theory showing how our algorithm, which we call Manifold Mixup, leads to a flattening of the per-class hidden representations, which can be seen as a compression of the information in the hidden states. The second article is related to the first and shows how interpolated examples can be used for semi-supervised learning. In addition to interpolating the input examples, the model’s interpolated predictions are used as targets for these examples. This improves results on standard benchmarks as well as classic 2D toy problems for semi-supervised learning. The third article studies how a recurrent neural network can be divided into multiple modules with different parameters and well separated hidden states, as well as a competition mechanism restricting updating of the hidden states to a subset of the most relevant modules on a specific time-step. This improves systematic generalization when the pattern distribution is changed between the training and evaluation phases. It also improves generalization in reinforcement learning. In the fourth article, we show that attention can be used to control the flow of information between successive layers in deep networks. This allows each layer to only process the subset of the previously computed layers’ outputs which are most relevant. This improves generalization on relational reasoning tasks as well as standard benchmark classification tasks. / Cette thèse explore la nature de la généralisation dans l’apprentissage en profondeur et plusieurs contextes dans lesquels elle échoue. En particulier, les réseaux de neurones profonds peuvent avoir du mal à se généraliser dans des contextes avec des données limitées, une supervision insuffisante, des dépendances à longue portée difficiles ou une structure et des sous-systèmes complexes. Cette thèse explore la nature de ces défis pour la généralisation en apprentissage profond et présente plusieurs algorithmes qui cherchent à relever ces défis. Dans le premier article, nous montrons comment l’entraînement avec des états cachés interpolés peut améliorer la généralisation et la calibration en apprentissage profond. Nous introduisons également une théorie montrant comment notre algorithme, que nous appelons Manifold Mixup, conduit à un aplatissement des représentations cachées par classe, ce qui peut être vu comme une compression de l’information dans les états cachés. Le deuxième article est lié au premier et montre comment des exemples interpolés peuvent être utilisés pour un apprentissage semi-supervisé. Outre l’interpolation des exemples d’entrée, les prédictions interpolées du modèle sont utilisées comme cibles pour ces exemples. Cela améliore les résultats sur les benchmarks standard ainsi que sur les problèmes de jouets 2D classiques pour l’apprentissage semi-supervisé. Le troisième article étudie comment un réseau de neurones récurrent peut être divisé en plusieurs modules avec des paramètres différents et des états cachés bien séparés, ainsi qu’un mécanisme de concurrence limitant la mise à jour des états cachés à un sous-ensemble des modules les plus pertinents sur un pas de temps spécifique. . Cela améliore la généralisation systématique lorsque la distribution des modèles est modifiée entre les phases de entraînement et d’évaluation. Il améliore également la généralisation dans l’apprentissage par renforcement. Dans le quatrième article, nous montrons que l’attention peut être utilisée pour contrôler le flux d’informations entre les couches successives des réseaux profonds. Cela permet à chaque couche de ne traiter que le sous-ensemble des sorties des couches précédemment calculées qui sont les plus pertinentes. Cela améliore la généralisation sur les tâches de raisonnement relationnel ainsi que sur les tâches de classification de référence standard.

Implicit communication is important in interaction because it plays a role in conveying the internal mental states of an individual. For example, emotional expressions that are shown through unintended facial gestures can communicate underlying affective states. People can infer mental states from implicit cues and have strong expectations of what those cues mean. This is true for human-human interactions, as well as human-robot interactions. A Normalizing flow model is used as a generative model that can produce facial gestures and head movements. The invertible nature of the Normalizing flow model makes it possible to manipulate attributes of the generated gestures. The model in this work is capable of generating facial expressions that look real and human-like. Furthermore, the model can manipulate the generated output to change the perceived affective state of the facial expressions. / Implicit kommunikation är viktig i interaktioner eftersom den spelar en roll för att förmedla individens inre mentala tillstånd. Till exempel kan känslomässiga uttryck som visas genom oavsiktliga ansiktsgester kommunicera underliggande affektiva tillstånd. Människor kan härleda mentala tillstånd från implicita ledtrådar och har starka förväntningar på vad dessa ledtrådar betyder. Detta gäller för interaktion mellan människor, liksom interaktion mellan människa och robot. En normaliserande flödesmodell används som en generativ modell som kan producera ansiktsgester och huvudrörelser. Den inverterbara naturen hos normaliseringsflödesmodellen gör det också möjligt att manipulera det genererade ansiktsuttrycken. Utgången manipuleras i två dimensioner som vanligtvis används för att beskriva affektivt tillstånd, valens och upphetsning. Modellen i detta arbete kan generera ansiktsuttryck som ser verkliga och mänskliga ut och kan manipuleras for att ändra det affektiva tillstånd.

Following the introduction of new laws and regulations to ensure data protection in GDPR and PIPEDA, interests in technologies to protect data privacy have increased. A promising research trajectory in this area is found in Generative Adversarial Networks (GAN), an architecture trained to produce data that reflects the statistical properties of its underlying dataset without compromising the integrity of the data subjects. Despite the technology’s young age, prior research has made significant progress in the generation process of so-called synthetic data, and the current models can generate images with high-quality. Due to the architecture’s success with images, it has been adapted to new domains, and this study examines its potential to synthesize financial tabular data. The study investigates a state-of-the-art model within tabular GANs, called CTGAN, together with two proposed ideas to enhance its generative ability. The results indicate that a modified training dynamic and a novel early stopping strategy improve the architecture’s capacity to synthesize data. The generated data presents realistic features with clear influences from its underlying dataset, and the inferred conclusions on subsequent analyses are similar to those based on the original data. Thus, the conclusion is that GANs has great potential to generate tabular data that can be considered a substitute for sensitive data, which could enable organizations to have more generous data sharing policies. / Med striktare förhållningsregler till hur data ska hanteras genom GDPR och PIPEDA har intresset för anonymiseringsmetoder för att censurera känslig data aktualliserats. En lovande teknik inom området återfinns i Generativa Motstridande Nätverk, en arkitektur som syftar till att generera data som återspeglar de statiska egenskaperna i dess underliggande dataset utan att äventyra datasubjektens integritet. Trots forskningsfältet unga ålder har man gjort stora framsteg i genereringsprocessen av så kallad syntetisk data, och numera finns det modeller som kan generera bilder av hög realistisk karaktär. Som ett steg framåt i forskningen har arkitekturen adopterats till nya domäner, och den här studien syftar till att undersöka dess förmåga att syntatisera finansiell tabelldata. I studien undersöks en framträdande modell inom forskningsfältet, CTGAN, tillsammans med två föreslagna idéer i syfte att förbättra dess generativa förmåga. Resultaten indikerar att en förändrad träningsdynamik och en ny optimeringsstrategi förbättrar arkitekturens förmåga att generera syntetisk data. Den genererade datan håller i sin tur hög kvalité med tydliga influenser från dess underliggande dataset, och resultat på efterföljande analyser mellan datakällorna är av jämförbar karaktär. Slutsatsen är således att GANs har stor potential att generera tabulär data som kan betrakatas som substitut till känslig data, vilket möjliggör för en mer frikostig delningspolitik av data inom organisationer.

Synthetic data provides a good alternative to real data when the latter is not sufficientor limited by privacy requirements. In spatio-temporal applications, generating syntheticdata is generally more complex due to the existence of both spatial and temporal dependencies.Recently, with the advent of deep generative modeling such as GenerativeAdversarial Networks (GAN), synthetic data generation has seen a lot of development andsuccess. This thesis uses a GAN model based on two Recurrent Neural Networks (RNN)as a generator and a discriminator to generate new trip data for transport vehicles, wherethe data is represented as a time series. This model is compared with a standalone RNNnetwork that does not have an adversarial counterpart. The result shows that the RNNmodel (without the adversarial counterpart) performed better than the GAN model dueto the difficulty that involves training and tuning GAN models.

The integration of artificial intelligence (AI) into biomedical research has given rise to new models and research topics in biomedicine. Whole-cell modeling aims to create a holistic understanding of the cell by integrating diverse data. One method of comprehension is the characterization and imitation of a system. Phenomenological cell models imitate cell structure and behavior based on, for example, images. Thus generative AI image models present one approach to developing such phenomenological models of cell systems. Diffusion models are a popular generative model class for image generation. Briefly, diffusion models consist of a forward and reverse diffusion process, where the forward process iteratively adds noise to an image and the reverse process learns to remove it. Image generation is achieved by sampling from noise and applying the learned reverse process. The generation may be conditioned to achieve a specific output. The diffusion process is computationally expensive to evaluate in pixel space. The latent diffusion model presents a solution by moving the diffusion process to the latent space of an autoencoder. A latent diffusion model has been trained to develop a phenomenological model of cells in mitosis. The aim is to identify spatial and temporal patterns in the dataset, consisting of fluorescence microscopy images of cells in mitosis, and condition the output of the latent diffusion model on labels associated with the data. The latent diffusion can generate images unconditionally and conditionally. The unconditionally generated images appear visually similar, but quantitative metrics suggest the potential for improvement. Qualitative analysis of the conditionally generated images indicates opportunities for enhancement. The analysis from the proposed method for objective assessment of conditionally generated images, feature extraction of images followed by dimension reduction using uniform manifold approximation and projection, concurs with the visual assessment. However, the quantitative metrics and the proposed method of conditional assessment rely upon InceptionV3 to extract features from the images. InceptionV3 has not been trained on biomedical images and thus the metrics and methods should not be overly relied upon. In general, there is a need for new assessment techniques suitable for non-class conditionally generated images that are unsuitable for evaluation using user studies. / Integrering av artificiell intelligens (AI) i biomedicinsk forskning har gett upphov till nya modeller och forskningsfrågor inom biomedicin. Helcellsmodellering syftar till att skapa ett kvantitativt perspektiv på cellbiologi och skapa holistisk kunskap om cellen. Ett system kan förstås genom karaktärisering och imitation. Generativ AI är ett tillvägagångssätt för att utveckla modeller som kan imitera och karaktärisera celler baserat på bilder. Diffusionsmodeller är en populär klass av generativa modeller för bildgenerering. Diffusionsmodeller består av en framåt- och bakåtdiffusionsprocess, där den framåtriktade processen iterativt lägger till brus i en bild och den bakåtriktade processen lär sig att ta bort det. Nya bilder genereras genom att tillämpa den inlärda bakåtriktade processen på en bild av brus. Generationen kan göras villkorlig för att forma bilden efter givna villkor. Den beräkningsintensiva diffusionsprocessen kan effektiviseras genom att introducera en "autoencoder" som flyttar diffusionsprocessen från pixelrummets stora dimension till det latenta rummet, som har en mindre dimension. Det utgör basen för en latent diffusionsmodell. För att utveckla en fenomenologisk modell av celler i mitos har en latent diffusionsmodell tränats på fluorescensmikroskopibilder på celler som genomgår mitos. Målet är att identifiera spatiala och temporala mönster i bilderna och skapa en modell som kan villkora bildgenerationen baserat på givna spatiala och temporala villkor associerade med bilderna. Latenta diffusionsmodeller kan skapa bilder både villkorligen och helt fritt från den underliggande datadistributionen. Den fria generationen av bilder resulterar i visuellt lika bilder men kvantitativa mått indikerar att modellen kan förbättras. Villkorligt genererade bilder håller inte samma visuella kvalité. Behovet av tekniker för att utvärdera villkorligt genererade bilder har identifierats och en metod har föreslagits. Metoden involverar att extrahera attribut från bilderna och reducera dimensionen av attributen för att visualisera de olika villkoren. Utvärderingen av de villkorligt genererade bilderna visar att den villkorliga generationen kan förbättras. Däremot beror metoden och de kvantitativa mått som beräknades för den fria generationen av bilder på ett neuralt nätverk som inte tränats på biomedicinska bilder. Därför bör resultaten tolkas med viss reservation.

Data augmentation is a technique that can be performed in various ways to improve the training of discriminative models. The recent developments in deep generative models offer new ways of augmenting existing data sets. In this thesis, a framework for augmenting annotated data sets with deep generative models is proposed together with a method for quantitatively evaluating the quality of the generated data sets. Using this framework, two data sets for pupil localization was generated with different generative models, including both well-established models and a novel model proposed for this purpose. The unique model was shown both qualitatively and quantitatively to generate the best data sets. A set of smaller experiments on standard data sets also revealed cases where this generative model could improve the performance of an existing discriminative model. The results indicate that generative models can be used to augment or replace existing data sets when training discriminative models. / Dataaugmentering är en teknik som kan utföras på flera sätt för att förbättra träningen av diskriminativa modeller. De senaste framgångarna inom djupa generativa modeller har öppnat upp nya sätt att augmentera existerande dataset. I detta arbete har ett ramverk för augmentering av annoterade dataset med hjälp av djupa generativa modeller föreslagits. Utöver detta så har en metod för kvantitativ evaulering av kvaliteten hos genererade data set tagits fram. Med hjälp av detta ramverk har två dataset för pupillokalisering genererats med olika generativa modeller. Både väletablerade modeller och en ny modell utvecklad för detta syfte har testats. Den unika modellen visades både kvalitativt och kvantitativt att den genererade de bästa dataseten. Ett antal mindre experiment på standardiserade dataset visade exempel på fall där denna generativa modell kunde förbättra prestandan hos en existerande diskriminativ modell. Resultaten indikerar att generativa modeller kan användas för att augmentera eller ersätta existerande dataset vid träning av diskriminativa modeller.