No description available.

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.

No description available.

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

Educational Digital Libraries (DLs) are complex information systems which are designed to support individuals' information needs and information seeking behavior. To have a broad impact on the communities in education and to serve for a long period, DLs need to structure and organize the resources in a way that facilitates the dissemination and the reuse of resources. Such a digital library should meet defined quality dimensions in the 5S (Societies, Scenarios, Spaces, Structures, Streams) framework - including completeness, consistency, efficiency, extensibility, and reliability - to ensure that a good quality DL is built. In this research, we addressed both external and internal quality aspects of DLs. For internal qualities, we focused on completeness and consistency of the collection, catalog, and repository. We developed an application pipeline to acquire user-generated computing-related resources from YouTube and SlideShare for an educational DL. We applied machine learning techniques to transfer what we learned from the ACM Digital Library dataset. We built classifiers to catalog resources according to the ACM Computing Classification System from the two new domains that were evaluated using Amazon Mechanical Turk. For external qualities, we focused on efficiency, scalability, and reliability in DL services. We proposed cloud-based designs and applications to ensure and improve these qualities in DL services using cloud computing. The experimental results show that our proposed methods are promising for enhancing and enriching an educational digital library. This work received support from ACM, as well as the National Science Foundation under Grant Numbers DUE-0836940, DUE-0937863, and DUE-0840719, and IMLS LG-71-16-0037-16. / Ph. D.

Le domaine de l’écologie des réseaux est encore limité dans sa capacité à faire des inférences mondiales à grande échelle. Ce défi est principalement dû à la difficulté d’échantillonnage des interactions sur le terrain, entraînant de nombreuses « lacunes » en ce qui concerne la couverture mondiale des données. Cette thèse adopte une approche « centrée sur les méthodes » de l’écologie des réseaux et se concentre sur l’idée de développer des outils pour aider à combler les lacunes en matière de données en présentant la prédiction comme une alternative accessible à l’échantillonnage sur le terrain et introduit deux « outils » différents qui sont prêts à poser des questions à l’échelle mondiale. Le chapitre 1 présente les outils que nous pouvons utiliser pour faire des prédictions de réseaux et est motivé par l’idée selon laquelle avoir la capacité de prédire les interactions entre les espèces grâce à l’utilisation d’outils de modélisation est impératif pour une compréhension plus globale des réseaux écologiques. Ce chapitre comprend une preuve de concept (dans laquelle nous montrons comment un simple modèle de réseau neuronal est capable de faire des prédictions précises sur les interactions entre espèces), une évaluation des défis et des opportunités associés à l’amélioration des prédictions d’interaction et une feuille de route conceptuelle concernant l’utilisation de modèles prédictifs pour les réseaux écologiques. Les chapitres 2 et 3 sont étroitement liés et se concentrent sur l’utilisation de l’intégration de graphiques pour la prédiction de réseau. Essentiellement, l’intégration de graphes nous permet de transformer un graphe (réseau) en un ensemble de vecteurs, qui capturent une propriété écologique du réseau et nous fournissent une abstraction simple mais puissante d’un réseau d’interaction et servent de moyen de maximiser les informations disponibles. dispo- nibles à partir des réseaux d’interactions d’espèces. Parce que l’intégration de graphes nous permet de « décoder » les informations au sein d’un réseau, elle est conçue comme un outil de prédiction de réseau, en particulier lorsqu’elle est utilisée dans un cadre d’apprentissage par transfert. Elle s’appuie sur l’idée que nous pouvons utiliser les connaissances acquises en résolvant un problème connu. et l’utiliser pour résoudre un problème étroitement lié. Ici, nous avons utilisé le métaweb européen (connu) pour prédire un métaweb pour les espèces canadiennes en fonction de leur parenté phylogénétique. Ce qui rend ce travail particulière- ment passionnant est que malgré le faible nombre d’espèces partagées entre ces deux régions, nous sommes capables de récupérer la plupart (91%) des interactions. Le chapitre 4 approfondit la réflexion sur la complexité des réseaux et les différentes ma- nières que nous pourrions choisir de définir la complexité. Plus spécifiquement, nous remet- tons en question les mesures structurelles plus traditionnelles de la complexité en présentant l’entropie SVD comme une mesure alternative de la complexité. Adopter une approche phy- sique pour définir la complexité nous permet de réfléchir aux informations contenues dans un réseau plutôt qu’à leurs propriétés émergentes. Il est intéressant de noter que l’entropie SVD révèle que les réseaux bipartites sont très complexes et ne sont pas nécessairement conformes à l’idée selon laquelle la complexité engendre la stabilité. Enfin, je présente le package Julia SpatialBoundaries.jl. Ce package permet à l’utili- sateur d’implémenter l’algorithme de wombling spatial pour des données disposées de manière uniforme ou aléatoire dans l’espace. Étant donné que l’algorithme de wombling spatial se concentre à la fois sur le gradient et sur la direction du changement pour un paysage donné, il peut être utilisé à la fois pour détecter les limites au sens traditionnel du terme ainsi que pour examiner de manière plus nuancée la direction des changements. Cette approche pourrait être un moyen bénéfique de réfléchir aux questions liées à la détection des limites des réseaux et à leur relation avec les limites environnementales. / The field of network ecology is still limited in its ability to make large-scale, global inferences. This challenge is primarily driven by the difficulty of sampling interactions in the field, leading to many ‘gaps’ with regards to global coverage of data. This thesis takes a ’methods-centric’ approach to network ecology and focuses on the idea of developing tools to help with filling in the the data gaps by presenting prediction as an accessible alternative to sampling in the field and introduces two different ’tools’ that are primed for asking questions at global scales. Chapter 1 maps out tools we can use to make network predictions and is driven by the idea that having the ability to predict interactions between species through the use of modelling tools is imperative for a more global understanding of ecological networks. This chapter includes a proof-of-concept (where we show how a simple neural network model is able to make accurate predictions about species interactions), an assessment of the challenges and opportunities associated with improving interaction predictions, and providing a conceptual roadmap concerned with the use of predictive models for ecological networks. Chapters 2 and 3 are closely intertwined and are focused on the use of graph embedding for network prediction. Essentially graph embedding allows us to transform a graph (net- work) into a set of vectors, which capture an ecological property of the network and provides us with a simple, yet powerful abstraction of an interaction network and serves as a way to maximise the available information available from species interaction networks. Because graph embedding allows us to ’decode’ the information within a network it is primed as a tool for network prediction, specifically when used in a transfer learning framework, this builds on the idea that we can take the knowledge gained from solving a known problem and using it to solve a closely related problem. Here we used the (known) European metaweb to predict a metaweb for Canadian species based on their phylogenetic relatedness. What makes this work particularly exciting is that despite the low number of species shared between these two regions we are able to recover most (91%) of interactions. Chapter 4 delves into thinking about the complexity of networks and the different ways we might choose to define complexity. More specifically we challenge the more traditional structural measures of complexity by presenting SVD entropy as an alternative measure of complexity. Taking a physical approach to defining complexity allows us to think about the information contained within a network as opposed to their emerging properties. Interest- ingly, SVD entropy reveals that bipartite networks are highly complex and do not necessarily conform to the idea that complexity begets stability. Finally, I present the Julia package SpatialBoundaries.jl. This package allows the user to implement the spatial wombling algorithm for data arranged uniformly or randomly across space. Because the spatial wombling algorithm focuses on both the gradient as well as the direction of change for the given landscape it can be used both for detecting boundaries in the traditional sense as well as a more nuanced look at at the direction of changes. This approach could be a beneficial way with which to think about questions which relate to boundary detection for networks and how these relate to environmental boundaries.

This dissertation comprises several studies addressing supervised learning problems where the supervision is imperfect. Firstly, we investigate the margin conditions in active learning. Active learning is characterized by its special mechanism where the learner can sample freely over the feature space and exploit mostly the limited labeling budget by querying the most informative labels. Our primary focus is to discern critical conditions under which certain active learning algorithms can outperform the optimal passive learning minimax rate. Within a non-parametric multi-class classification framework,our results reveal that the uniqueness of Bayes labels across the feature space serves as the pivotal determinant for the superiority of active learning over passive learning. Secondly, we study the estimation of central mean subspace (CMS), and its application in transfer learning. We show that a fast parametric convergence rate is achievable via estimating the expected smoothed gradient outer product, for a general class of covariate distribution that admits Gaussian or heavier distributions. When the link function is a polynomial with a degree of at most r and the covariates follow the standard Gaussian, we show that the prefactor depends on the ambient dimension d as d^r. Furthermore, we show that under a transfer learning setting, an oracle rate of prediction error as if the CMS is known is achievable, when the source training data is abundant. Finally, we present an innovative application involving the utilization of weak (noisy) labels for addressing an Individual Tree Crown (ITC) segmentation challenge. Here, the objective is to delineate individual tree crowns within a 3D LiDAR scan of tropical forests, with only 2D noisy manual delineations of crowns on RGB images available as a source of weak supervision. We propose a refinement algorithm designed to enhance the performance of existing unsupervised learning methodologies for the ITC segmentation problem.

<p dir="ltr">Broadly, the focus of this work is on efficient statistical estimation and optimization of data arising from experimental data, particularly motivated by nanomanufacturing experiments on the material tellurene. Tellurene is a novel material for transistors with reliable attributes that enhance the performance of electronics (e.g., nanochip). As a solution-grown product, two-dimensional (2D) tellurene can be manufactured through a scalable process at a low cost. There are three main throughlines to this work, data augmentation, optimization, and equality constraint, and three distinct methodological projects, each of which addresses a subset of these throughlines. For the first project, I apply transfer learning in the analysis of data from a new tellurene experiment (process B) using the established linear regression model from a prior experiment (process A) from a similar study to combine the information from both experiments. The key of this approach is to incorporate the total equivalent amounts (TEA) of a lurking variable (experimental process changes) in terms of an observed (base) factor that appears in both experimental designs into the prespecified linear regression model. The results of the experimental data are presented including the optimal PVP chain length for scaling up production through a larger autoclave size. For the second project, I develop a multi-armed bandit Bayesian optimization (BO) approach to incorporate the equality constraint that comes from a mixture experiment on tellurium nanoproduct and account for factors with categorical levels. A more complex optimization approach was necessitated by the experimenters’ use of a neural network regression model to estimate the response surface. Results are presented on synthetic data to validate the ability of BO to recover the optimal response and its efficiency is compared to Monte Carlo random sampling to understand the level of experimental design complexity at which BO begins to pay off. The third project examines the potential enhancement of parameter estimation by utilizing synthetic data generated through Generative Adversarial Networks (GANs) to augment experimental data coming from a mixture experiment with a small to moderate number of runs. Transfer learning shows high promise for aiding in tellurene experiments, BO’s value increases with the complexity of the experiment, and GANs performed poorly on smaller experiments introducing bias to parameter estimates.</p>

This report presents a deep learning based approach for segmenting and characterizing tumor cell deaths using images provided by the Önfelt lab, which contain NK cells and HL60 leukemia cells. We explore the efficiency of convolutional neural networks (CNNs) in distinguishing between live and dead tumor cells, as well as different classes of cell death. Three CNN architectures: MobileNetV2, ResNet-18, and ResNet-50 were employed, utilizing transfer learning to optimize performance given the limited size of available datasets. The networks were trained using two loss functions: weighted cross-entropy and generalized dice loss and two optimizers: Adaptive moment estimation (Adam) and stochastic gradient descent with momentum (SGDM), with performance evaluations based on metrics such as mean accuracy, intersection over union (IoU), and BF score. Our results indicate that MobileNetV2 with cross-entropy loss and the Adam optimizer outperformed other configurations, demonstrating high mean accuracy. Challenges such as class imbalance, annotation bias, and dataset limitations are discussed, alongside potential future directions to enhance model robustness and accuracy. The successful training of networks capable of classifying all identified types of cell death, demonstrates the potential for a deep learning approach to identify different types of cell deaths as a tool for analyzing immunotherapeutic strategies and enhance understanding of NK cell behaviors in cancer treatment.

Models based around the transformer architecture have become one of the most prominent for solving a multitude of natural language processing (NLP)tasks since its introduction in 2017. However, much research related to the transformer model has focused primarily on achieving high performance and many problems remain unsolved. Two of the most prominent currently are the lack of high performing non-English pre-trained models, and the limited number of words most trained models can incorporate for their context. Solving these problems would make NLP models more suitable for real-world applications, improving information retrieval, reading comprehension, and more. All previous research has focused on incorporating long-context for English language models. This thesis investigates the cross-lingual transferability between languages when only training for long-context in English. Training long-context models in English only could make long-context in low-resource languages, such as Swedish, more accessible since it is hard to find such data in most languages and costly to train for each language. This could become an efficient method for creating long-context models in other languages without the need for such data in all languages or pre-training from scratch. We extend the models’ context using the training scheme of the Longformer architecture and fine-tune on a question-answering task in several languages. Our evaluation could not satisfactorily confirm nor deny if transferring long-term context is possible for low-resource languages. We believe that using datasets that require long-context reasoning, such as a multilingual TriviaQAdataset, could demonstrate our hypothesis’s validity.