<pre><pre><p> </p><p>Recent progress in machine learning has been mainly due to the availability of large amounts of annotated data used for training complex models with deep architectures. Annotating this training data becomes burdensome and creates a major bottleneck in maintaining machine-learning databases. Moreover, these trained models fail to generalize to new categories or new varieties of the same categories. This is because new categories or new varieties have data distribution different from the training data distribution. To tackle these problems, this thesis proposes to develop a family of transfer-learning techniques that can deal with different training (source) and testing (target) distributions with the assumption that the availability of annotated data is limited in the testing domain. This is done by using the auxiliary data-abundant source domain from which useful knowledge is transferred that can be applied to data-scarce target domain. This transferable knowledge serves as a prior that biases target-domain predictions and prevents the target-domain model from overfitting. Specifically, we explore structural priors that encode relational knowledge between different data entities, which provides more informative bias than traditional priors. The choice of the structural prior depends on the information availability and the similarity between the two domains. Depending on the domain similarity and the information availability, we divide the transfer learning problem into four major categories and propose different structural priors to solve each of these sub-problems.</p><p> </p><p>This thesis first focuses on the unsupervised-domain-adaptation problem, where we propose to minimize domain discrepancy by transforming labeled source-domain data to be close to unlabeled target-domain data. For this problem, the categories remain the same across the two domains and hence we assume that the structural relationship between the source-domain samples is carried over to the target domain. Thus, graph or hyper-graph is constructed as the structural prior from both domains and a graph/hyper-graph matching formulation is used to transform samples in the source domain to be closer to samples in the target domain. An efficient optimization scheme is then proposed to tackle the time and memory inefficiencies associated with the matching problem. The few-shot learning problem is studied next, where we propose to transfer knowledge from source-domain categories containing abundantly labeled data to novel categories in the target domain that contains only few labeled data. The knowledge transfer biases the novel category predictions and prevents the model from overfitting. The knowledge is encoded using a neural-network-based prior that transforms a data sample to its corresponding class prototype. This neural network is trained from the source-domain data and applied to the target-domain data, where it transforms the few-shot samples to the novel-class prototypes for better recognition performance. The few-shot learning problem is then extended to the situation, where we do not have access to the source-domain data but only have access to the source-domain class prototypes. In this limited information setting, parametric neural-network-based priors would overfit to the source-class prototypes and hence we seek a non-parametric-based prior using manifolds. A piecewise linear manifold is used as a structural prior to fit the source-domain-class prototypes. This structure is extended to the target domain, where the novel-class prototypes are found by projecting the few-shot samples onto the manifold. Finally, the zero-shot learning problem is addressed, which is an extreme case of the few-shot learning problem where we do not have any labeled data in the target domain. However, we have high-level information for both the source and target domain categories in the form of semantic descriptors. We learn the relation between the sample space and the semantic space, using a regularized neural network so that classification of the novel categories can be carried out in a common representation space. This same neural network is then used in the target domain to relate the two spaces. In case we want to generate data for the novel categories in the target domain, we can use a constrained generative adversarial network instead of a traditional neural network. Thus, we use structural priors like graphs, neural networks and manifolds to relate various data entities like samples, prototypes and semantics for these different transfer learning sub-problems. We explore additional post-processing steps like pseudo-labeling, domain adaptation and calibration and enforce algorithmic and architectural constraints to further improve recognition performance. Experimental results on standard transfer learning image recognition datasets produced competitive results with respect to previous work. Further experimentation and analyses of these methods provided better understanding of machine learning as well.</p><p> </p></pre></pre>

Orientador: José Augusto Chaves Guimarães / Banca: João Batista Ernesto de Moraes / Banca: Eduardo Ismael Murguia Marañon / Banca: Juan Carlos Fernández-Molina / Banca: Miriam Figueiredo Vieira da Cunha / Resumo: Os estudos sobre a ética na Organização e Representação do Conhecimento, especialmente no Tratamento Temático da Informação, têm colaborado para sedimentar os referenciais teóricos e metodológicos da Ciência da Informação, que se justificam pelo pressuposto da inclusão social que, enquanto um metavalor, se situa entre o preconceito social e o proselitismo, formando um cenário onde três universos axiológicos convivem: o do documento ou informação, o do usuário e o do bibliotecário. Por isso, a indexação está ligada a uma dimensão ética porque deve preocupar-se com sua confiabilidade e utilidade em relação a determinadas comunidades discursivas ou domínios específicos. Nesse sentido, propõe-se, por meio de uma pesquisa exploratória e documental, com características qualitativas e indutivas, identificar a máxima especificidade terminológica que linguagens de indexação brasileiras permitem para termos relativos à homossexualidade masculina, analisando como corpus investigativo os termos atribuídos aos artigos científicos publicados na Journal of Homosexuality, Sexualities e Journal of Gay & Lesbian Mental Health, entre os anos de 2005 a 2009. Do cotejo e análise dos termos e das linguagens de indexação brasileiras verifica-se uma aproximação de significados no contexto brasileiro, imprecisão terminológica, com indícios de preconceitos disseminados através do 'politicamente correto', representação inadequada da temática e a presença de figuras de linguagem / Abstract: The studies on ethics in Knowledge Organization and Representation, especially in the Subject Approach to Information, have collaborated to establish the theoretical and methodological aspects of Information Science which are justified by the assumption of social inclusion, as a metavalue, it situated itself between social prejudice and proselytize, creating a situation where three axiological universes coexist: the document or information, the user and the librarian. Therefore, the indexing is linked to an ethical dimension because it must concern itself with its reliability and usefulness in certain discourse communities or specific domains. In this direction, it is proposed through an exploratory and documental research with qualitative and inductive characteristics to identify the maximum specific terminological that Brazilian indexing languages allow for terms relating to male homosexuality, analyzing like investigative corpus the terms assigned to papers published in the Journal of Homosexuality, Sexualities and Journal of Gay & Lesbian Mental Health, between the years 2005 to 2009. From confrontation and analysis of terms and the Brazilian indexing languages there is an approximation of meaning in the Brazilian context, imprecision in the terminology, with indications of prejudices disseminate by 'politically correct', the biased representation of the thematic and the presence of figures of speech / Doutor

For a long time, there has been little research on dynamic social networks. However, in recent years, there has been much more focus on this field and many techniques for analyzing temporal aspects of social networks were proposed. In this work, we studied a dynamic social network based on data retrieved from the Commercial Register. This registry contains information about all economic entities that operate in the Czech Republic, including people who hold functions in entities and their addresses of living. We applied several data analysis techniques including community tracing, clustering, and methods for identifying key actors to find important entities and individuals in the social network and inspect their changes over time. 1

The goal of this master thesis was to propose a suitable strategy to detect and classify objects of interest in mammogram images. A part of this goal was to implement an experimentation framework, that will be used for data preparation, model training and comparison. Patch and full-image versions of the dataset were used in the analysis. Initialisation with weights that were pretrained on the images from other domain improved classifier performance. ResNet-34 had better AUC scores on the test set that ResNet-18. Semi-supervised training using entropy minimisation has no significant improvement over the supervised one. The thesis includes the visualisation of the network predictions and the analysis of the knowledge representation of the classier. The achieved results for a patch version of the dataset are comparable to the results of another article that utilised the same test set. For a full-image dataset the results of the training were suboptimal. 1

No description available.

Ontologien spielen eine zentrale Rolle für die formalisierte Repräsentation von Wissen und Informationen sowie für die Infrastruktur des sogenannten semantic web. Trotz früherer Initiativen der Bibliotheken und Gedächtnisinstitutionen hat sich die deutschsprachige Musikwissenschaft insgesamt nur sehr zögerlich dem Thema genähert. Im Rahmen einer Bestandsaufnahme werden neben der Erläuterung grundlegender Konzepte, Herausforderungen und Herangehensweisen bei der Modellierung von Ontologien daher auch vielversprechende Modelle und bereits erprobte Anwendungsbeispiele für eine ‚semantische‘ digitale Musikwissenschaft identifiziert. / Ontologies play a crucial role for the formalised representation of knowledge and information as well as for the infrastructure of the semantic web. Despite early initiatives that were driven by libraries and memory institutions, German musicology as a whole has turned very slowly to the subject. In an overview the author addresses basic concepts, challenges, and approaches for ontology design and identifies models and use cases with promising applications for a ‚semantic‘ digital musicology.

<div>In the past fifty years, Deep Neural Networks (DNNs) have evolved greatly from a single perceptron to complex multi-layered networks with non-linear activation functions. Today, they form the backbone of Artificial Intelligence, with a diverse application landscape, such as smart assistants, wearables, targeted marketing, autonomous vehicles, etc. The design of DNNs continues to change, as we push its abilities to perform more human-like tasks at an industrial scale.</div><div><br></div><div>Multi-task learning and knowledge sharing are essential to human-like learning. Humans progressively acquire knowledge throughout their life, and they do so by remembering, and modifying prior skills for new tasks. In our first work, we investigate the representations learned by Spiking Neural Networks (SNNs), and how to share this knowledge across tasks. Our prior task was MNIST image generation using a spiking autoencoder. We combined the generative half of the autoencoder with a spiking audio-decoder for our new task, i.e audio-to-image conversion of utterances of digits to their corresponding images. We show that objects of different modalities carrying the same meaning can be mapped into a shared latent space comprised of spatio-temporal spike maps, and one can transfer prior skills, in this case, image generation, from one task to another, in a purely Spiking domain. Next, we propose Tree-CNN, an adaptive hierarchical network structure composed of Deep Convolutional Neural Networks(DCNNs) that can grow and learn as new data becomes available. The network organizes the incrementally available data into feature-driven super-classes and improves upon existing hierarchical CNN models by adding the capability of self-growth. </div><div><br></div><div>While the above works focused solely on algorithmic design, the underlying hardware determines the efficiency of model implementation. Currently, neural networks are implemented in CMOS based digital hardware such as GPUs and CPUs. However, the saturating scaling trend of CMOS has garnered great interest in Non-Volatile Memory (NVM) technologies such as Spintronics and RRAM. However, most emerging technologies have inherent reliability issues, such as stochasticity and non-linear device characteristics. Inspired by the recent works in spin-based stochastic neurons, we studied the algorithmic impact of designing a neural network using stochastic activations. We trained VGG-like networks on CIFAR-10/100 with 4 different binary activations and analyzed the trade-off between deterministic and stochastic activations. </div><div><br></div><div>NVM-based crossbars further promise fast and energy-efficient in-situ matrix-vector multiplications (MVM). However, the analog nature of computing in these NVM crossbars introduces approximations in the MVM operations, resulting in deviations from ideal output values. We first studied the impact of these non-idealities on the performance of vanilla DNNs under adversarial circumstances, and we observed that the non-ideal behavior interferes with the computation of the exact gradient of the model, which is required for adversarial image generation. In a non-adaptive attack, where the attacker is unaware of the analog hardware, analog computing offered varying degree of intrinsic robustness under all attack scenarios - Transfer, Black Box, and White Box attacks. We also demonstrated ``Hardware-in-Loop" adaptive attacks that circumvent this robustness by utilizing the knowledge of the NVM model.</div><div><br></div><div>Next, we explored the design of robust DNNs through the amalgamation of adversarial training and the intrinsic robustness offered by NVM crossbar based analog hardware. We studied the noise stability of such networks on unperturbed inputs and observed that internal activations of adversarially trained networks have lower Signal-to-Noise Ratio (SNR), and are sensitive to noise than vanilla networks. As a result, they suffer significantly higher performance degradation due to the non-ideal computations, on an average 2x accuracy drop. On the other hand, for adversarial images, the same networks displayed a 5-10% gain in robust accuracy due to the underlying NVM crossbar when the attack epsilon (the degree of input perturbations) was greater than the epsilon of the adversarial training. Our results indicate that implementing adversarially trained networks on analog hardware requires careful calibration between hardware non-idealities and training epsilon to achieve optimum robustness and performance.</div>

<p>Empowering machines to understand our physical world should go beyond models with only natural language and models with only vision. Vision and language is a growing field of study that attempts to bridge the gap between natural language processing and computer vision communities by enabling models to learn visually grounded language. However, as an increasing number of pre-trained visual linguistic models focus on the alignment between visual regions and natural language, it is difficult to claim that these models capture certain properties of objects in their latent space, such as size. Inspired by recent trends in prompt learning, this study will design a prompt learning framework for two visual linguistic models, ViLBERT and ViLT, and use different manually crafted prompt templates to evaluate the consistency of performance of these models in comparing the size of objects. The results of this study showed that ViLT is more consistent in prediction accuracy for the given task with six pairs of objects under four prompt designs. However, the overall prediction accuracy is lower than the expectation on this object size comparison task; even the better model in this study, ViLT, has only 16 out of 24 cases better than the proposed random chance baseline. As this study is a preliminary study to explore the potential of pre-trained visual linguistic models on object size comparison, there are many directions for future work, such as investigating more models, choosing more object pairs, and trying different methods for feature engineering and prompt engineering.</p>

<p>Development of complex aerospace systems often takes decades of research and testing. High  performing propellants are important to the success of rocket propulsion systems. Development  and testing of new propellants can be expensive and dangerous. Full scale tests are often required  to understand the performance of new propellants. Many industries have started using data science  tools to learn from previous work and conduct smarter tests. Material scientists have started using  these tools to speed up the development of new materials. These data science tools can be used to  speed up the development and design better propellants. I approach the development of new solid  propellants through two steps: Prediction of delivered performance from available literature tests,  prediction of ideal performance using physics-based models. Random Forest models are used to  correlate the ideal performance to delivered performance of a propellant based on the composition  and motor properties. I use Parsimonious Neural Networks (PNNs) to learn interpretable models  for the ideal performance of propellants. I find that the available open literature data is too biased  for the models to learn from and discover families of interpretable models to predict the ideal  performance of propellants. </p>

Ce travail de thèse se situe dans le cadre du projet Terra Dynamica visant à peupler une ville virtuelle avec des agents qui simulent des piétons et des véhicules. L’objectif de notre travail est de rendre l’environnement compréhensible par les agents de la simulation afin qu’ils puissent exhiber des comportements crédibles. Les premiers travaux qui ont été proposés pour la modélisation sémantique des environnements virtuels gardent toujours un lien de dépendance avec la représentation graphique pré-existante de l’environnement. Cependant, l’information sémantique représentée dans ce genre d’approches est difficilement exploitable par les agents pour effectuer des procédures de raisonnement complexes en dehors des algorithmes de navigation. Nous présentons dans cette thèse un modèle de représentation de la sémantique de l’environnement qui fournit aux agents des données sur l’utilisation des objets de l’environnement pour permettre au mécanisme d’aide à la décision de produire des comportements crédibles. Par ailleurs, en réponse à des contraintes inhérentes à la simulation urbaine, notre approche est capable de traiter un grand nombre d’agents, en temps réel. Notre modèle est basé sur le principe que les objets de l’environnement proposent des services permettant de réaliser les actions avec différentes qualités. Nous avons donc représenté les informations sémantiques des objets liées à leur utilisation sous forme de services dans une ontologie de services. Nous avons utilisé cette ontologie de services pour calculer une qualité de service QoS qui nous permet de trier les différents objets permettant de réaliser une même action. Ainsi, nous pouvons comparer entre les services proposés par les objets pour proposer aux agents les meilleurs objets leur permettant de réaliser leurs actions afin d’acquérir une crédibilité comportementale. Afin d’évaluer l’impact de notre modèle sur la crédibilité des comportements produits, nous avons défini un protocole d’évaluation dédié aux modèles de représentation de la sémantique dans les environnements. Dans ce protocole, des observateurs doivent évaluer le caractère crédible des comportements produits par le simulateur à partir d’un modèle sémantique de l’environnement. Grâce à cette évaluation, nous montrons que notre modèle permet de simuler des agents dont le comportement est jugé comme crédible par des observateurs humains. Nous présentons également une évaluation qualitative de la capacité de notre modèle de passer à l’échelle et de répondre aux contraintes d’une simulation temps-réel. Cette évaluation, nous a permis de montrer que les caractéristiques de l’architecture de notre modèle nous permettent de répondre en un temps raisonnable aux demandes d’un grand nombre d’agents. / This work is part of the Terra Dynamica project whose objective was to populate a virtual city with agents that simulate pedestrians and vehicles. The aim of our work is to make agents which understand their environment so they can produce credible behaviors The first proposed solutions for the semantic modeling of virtual environments still keep a link with the pre-existing graphic representation of the environment. However, the semantic information represented in this kind of approach is difficult to use by the agents to perform complex reasoning procedures outside the navigation algorithms. In this thesis we present a semantic representation model of the environment that provides the agents with data on the use of environmental objects in order to allow the decision mechanism to produce credible behaviors. Furthermore, in response to the constraints that are inherent to the urban simulation, our approach is capable of handling a large number of agents in real time. Our model is based on the principle that environmental objects provide services for performing actions with different qualities. We have therefore represented the semantic information of the objects related to their use, as services in an ontology of services. We used this ontology of services to calculate a QoS which allows us to sort the different objects which all perform the same action. Thus, we can compare between the services offered by different objects in order to provide the agents with the best objects that allow them to carry out their actions and exhibit behavioral credibility. To assess the impact of our model on the credibility of the produced behaviors, we defined an evaluation protocol for the semantic representation of virtual environment models. In this protocol, observers must assess the credibility of behaviors produced by the simulator using a semantic model of the environment. Through this evaluation, we show that our model can simulate agents whose behavior is deemed credible by human observers. We also present a qualitative assessment of the ability of our model to scale and meet the constraints of a real-time simulation. This evaluation allowed us to show that the characteristics of the architecture of our model allow us to respond in a reasonable amount of time to requests from a large number of agents.