Thèses sur le sujet « Neural border »

The vertebrate nervous system is extremely complex and contains a wide diversity of cell types. The formation of a functional nervous system requires the differential specification of progenitor cells at the right time and place. The generation of many different types of neurons along the rostro-caudal axis of the CNS begins with the initial specification of a few progenitor domains. This initial coarse pattern is refined by so-called secondary organizers arising at boundaries between these domains. The Isthmic Organizer (IsO) is a secondary organizer located at the boundary between the midbrain and the hindbrain. Although the function and maintenance of the IsO are well understood, the processes underlying its initial specification have remained elusive. In the present work we provide evidence that convergent Wnt and FGF signals initiate the specification of the IsO during late gastrulation as part of the neural caudalization process. The initial step in the generation of the nervous system is the division of the embryonic ectoderm into three cell populations: neural cells giving rise to the CNS, neural plate border cells giving rise to the peripheral nervous system, and epidermal cells giving rise to the outer layer of the skin. While the choice between neural and epidermal fate has been well studied, the mechanism by which neural plate border cells are generated is less well understood. At rostral levels of the neuraxis, the neural plate border gives rise to the olfactory and lens placodes, thickenings of the surface ectoderm from which sensory organs are derived. More caudally, the neural plate border generates neural crest cells, a transient population that migrates extensively and contributes to neurons and glia of the peripheral nervous system. How the early patterning of the central and peripheral nervous systems are coordinated has remained poorly understood. Here we show that the generation of neural plate border cells is initiated at the late blastula stage and involves two phases. During the first phase, neural plate border cells are exposed to Wnt signals in the absence of BMP signals. Simultaneous exposure to Wnt and BMP signals at this early stage leads to epidermal induction. Wnt signals induce expression of Bmp4, thereby regulating the sequential exposure of cells to Wnt and BMP signals. During the second phase, at the late gastrula stage, BMP signals play an instructive role to specify neural plate border cells of either placodal or neural crest character depending on the status of Wnt signaling. At this stage, Wnt signals promote caudal character simultaneously in the neural plate border and in the neural ectoderm. Thus, the choice between epidermal and neural plate border specification is mediated by an interplay of Wnt and BMP signals that represents a novel mechanism involving temporal control of BMP activity by Wnt signals. Moreover, the early development of the central and peripheral nervous systems are coordinated by simultaneous caudalization by Wnt signals.

Por ser não-invasiva e de baixo custo, a Ecocardiografia tem se tornado uma técnica de diagnóstico muito utilizada para a determinação dos volumes sistólicos e diastólicos do ventrículo esquerdo a fim de se calcular, indiretamente, o volume de ejeção do ventrículo esquerdo, a razão de contração muscular das cavidades cardíacas, a fração de ejeção regional e global, a espessura do miocárdio e a massa ventricular. Para isso, torna-se necessária a detecção das bordas endocárdicas do ventrículo esquerdo, o que é dificultada pelo fato da imagem de Ecocardiografia possuir ruídos que prejudicam sua definição. Apesar de haver várias técnicas de segmentação de imagem, este trabalho propõe detectar as bordas do ventrículo esquerdo de imagens ecocardiográficas utilizando uma rede neural artificial para reconhecer padrões de bordas. A fim de acelerar o processo e facilitar o processamento, uma área retangular centrada dentro da janela acústica do paciente é determinada pelo operador com o uso do \'mouse\' na qual serão realizadas todas as análises e reconhecimentos de borda pela rede neural. Após a marcação dos pontos reconhecidos pela rede neural como bordas, utilizam-se técnicas de gradientes e contorno móvel para se conectar os pontos de maior probabilidade e traçar a borda do ventrículo esquerdo. Esta técnica mostrou-se eficaz quando comparados com as bordas traçadas pelo especialista, sendo um fator importante a prática do operador ao escolher adequadamente a área a ser analisada. Após treinamento com 50 amostras de padrões de \"borda\" e 10 amostras de padrões de \"não borda\", a técnica foi testada em 108 imagens, alcançando resultados com boa precisão e rapidez quando comparamos os resultados na determinação da área do ventrículo esquerdo com outras técnicas citadas na literatura nacional e internacional.
Being non-invasive and having low cost, the echocardiography has been largely applied as diagnostic technique for left ventricle systolic and diastolic volumes determination that indirectly are used to calculate the left ventricle ejection volume, the cardiac cavities muscular contraction, the regional and global ejection fraction, the myocardial thickness, the ventricular mass, etc. For this reason, the detection of the left ventricle endocardial borders become necessary, but hampered by the noise that impairs the echocardiography images definition. In spite of having many image segmentation techniques, this work intend to detect the borders of left ventricle on echocardiography images by using a artificial neural network to recognize border patterns. To accelerate the process and facilitate the procedure, the operator uses the mouse to define a rectangular region inside the acoustic window of the pacient where all analyses and border recognitions will be accomplished. After labeling the recognized points as \'border\', gradient techniques and mobile boundary are used to connect the points of greater probability and delineate the left ventricle border. This technique has proved to be efficient when compared to the borders traced by the specialist. The ability of the operator is important in choosing of the region to be analyzed. After training with 50 samples of \"border\" pattern and 10 samples of \"no-border\" pattern, this technique was tested on 108 images, achieving good results on precision and velocitiy when we compared the calculated left ventricle area with the results of other techniques published on national and international literature.

Thesis (Ph.D. in Neuroscience) -- University of Colorado Denver, 2008.
Includes bibliographical references (leaves 112-120). Free to UCD affiliates. Online version available via ProQuest Digital Dissertations;

The size and speed of computer networks continue to expand at a rapid pace, as do the corresponding errors, failures, and faults inherent within such extensive networks. This thesis introduces a novel approach to interface Border Gateway Protocol (BGP) computer networks with neural networks to learn the precursor connectivity patterns that emerge prior to a node failure. Details of the design and construction of a framework that utilizes neural networks to learn and monitor BGP connection states as a means of detecting and predicting BGP peer node failure are presented. Moreover, this framework is used to monitor a BGP network and a suite of tests are conducted to establish that this neural network approach as a viable strategy for predicting BGP peer node failure. For all performed experiments both of the proposed neural network architectures succeed in memorizing and utilizing the network connectivity patterns. Lastly, a discussion of this framework's generic design is presented to acknowledge how other types of networks and alternate machine learning techniques can be accommodated with relative ease.

It has been proposed that as animals explore their environment they build and maintain a cognitive map, an internal representation of their surroundings (Tolman, 1948). We tested this hypothesis using a task designed to assess the ability of rats to make a spatial inference (take a novel shortcut)(Roberts et al., 2007). Our findings suggest that rats are unable to make a spontaneous spatial inference. Furthermore, they bear similarities to experiments which have been similarly unable to replicate or support Tolman’s (1948) findings. An inability to take novel shortcuts suggests that rats do not possess a cognitive map (Bennett, 1996). However, we found evidence of alternative learning strategies, such as latent learning (Tolman & Honzik, 1930b) , which suggest that rats may still be building such a representation, although it does not appear they are able to utilise this information to make complex spatial computations. Neurons found in the hippocampus show remarkable spatial modulation of their firing rate and have been suggested as a possible neural substrate for a cognitive map (O'Keefe & Nadel, 1978). However, the firing of these place cells often appears to be modulated by features of an animal’s behaviour (Ainge, Tamosiunaite, et al., 2007; Wood, Dudchenko, Robitsek, & Eichenbaum, 2000). For instance, previous experiments have demonstrated that the firing rate of place fields in the start box of some mazes are predictive of the animal’s final destination (Ainge, Tamosiunaite, et al., 2007; Ferbinteanu & Shapiro, 2003). We sought to understand whether this prospective firing is in fact related to the goal the rat is planning to navigate to or the route the rat is planning to take. Our results provide strong evidence for the latter, suggesting that rats may not be aware of the location of specific goals and may not be aware of their environment in the form of a contiguous map. However, we also found behavioural evidence that rats are aware of specific goal locations, suggesting that place cells in the hippocampus may not be responsible for this representation and that it may reside elsewhere (Hok, Chah, Save, & Poucet, 2013). Unlike their typical activity in an open field, place cells often have multiple place fields in geometrically similar areas of a multicompartment environment (Derdikman et al., 2009; Spiers et al., 2013). For example, Spiers et al. (2013) found that in an environment composed of four parallel compartments, place cells often fired similarly in multiple compartments, despite the active movement of the rat between them. We were able to replicate this phenomenon, furthermore, we were also able to show that if the compartments are arranged in a radial configuration this repetitive firing does not occur as frequently. We suggest that this place field repetition is driven by inputs from Boundary Vector Cells (BVCs) in neighbouring brain regions which are in turn greatly modulated by inputs from the head direction system. This is supported by a novel BVC model of place cell firing which predicts our observed results accurately. If place cells form the neural basis of a cognitive map one would predict spatial learning to be difficult in an environment where repetitive firing is observed frequently (Spiers et al., 2013). We tested this hypothesis by training animals on an odour discrimination task in the maze environments described above. We found that rats trained in the parallel version of the task were significantly impaired when compared to the radial version. These results support the hypothesis that place cells form the neural basis of a cognitive map; in environments where it is difficult to discriminate compartments based on the firing of place cells, rats find it similarly difficult to discriminate these compartments as shown by their behaviour. The experiments reported here are discussed in terms of a cognitive map, the likelihood that such a construct exists and the possibility that place cells form the neural basis of such a representation. Although the results of our experiments could be interpreted as evidence that animals do not possess a cognitive map, ultimately they suggest that animals do have a cognitive map and that place cells form a more than adequate substrate for this representation.

Le but des recherches du laboratoire de génétique du développement est de mieux comprendre les mécanismes moléculaires qui contrôlent le développement neural des vertébrés. C’est la raison pour la quelle, j’ai identifié deux EST (BC071005 et BC077938) spécifiques de l’expression génique chez le Xenopus laevis. Sur base de la littérature, ces deux gènes présentent des profils d’expression intéressants, caractéristiques des gènes impliqués dans la neurogenèse.

Le premier clone d’ADNc code pour l’homologue du facteur de transcription Homez, contenant trois homéodomaines et deux motifs leucine zipper et dont la fonction est inconnue. Mes résultats ont montré que chez l’embryon de xénope au stade neurula, Homez est exprimé préférentiellement dans la plaque neurale, l’expression la plus forte étant détectée dans les domaines où les neurones primaires apparaissent. Plus tard, Homez est détecté dans le tube neural dans des cellules neurales postmitotiques en cours de différenciation. En accord avec ce profil d’expression, j’ai observé que Homez est régulé positivement par l’atténuation des signaux BMPs et par le facteur proneural Ngnr1 et négativement par la voie Notch. Bien que le facteur Homez ne soit pas suffisant pour induire une expression ectopique de marqueurs neuronaux dans l’embryon de xénope, j’ai pu montrer, en utilisant une approche de morpholino antisens, que celui-ci est requis en aval du facteur Ngnr1 pour la différenciation des précurseurs neuraux en neurones primaires.

Le deuxième clone code pour l’homologue du facteur CBFβ qui s’associe avec une famille de protéines CBFα1-3/Aml1-3/Runx1-3 pour former un complexe hétérodimérique liant l’ADN. Alors que chez la souris, les facteurs Runx1 et Runx3 jouent un rôle important dans la neurogenèse dans les ganglions spinaux et que chez le xénope, Runx1 est requis pour la formation des neurones Rohon-Beard, le rôle de CBFβ au cours du développement du système nerveux est actuellement mal connu. Mes résultats ont montré que chez l’embryon de xénope au stade neurula, CBFβ est coexprimé avec les facteurs Runx1-3 en bordure de la plaque neurale, mais de manière plus étendue et plus précoce. Comme attendu pour un marqueur de la bordure de la plaque neurale, j’ai observé que l’expression de CBFβ est régulée par les signaux BMP, Wnt, FGF et Notch. De manière intéressante, son expression est induite par les facteurs proneuraux alors que celle de Runx1 est inhibée. Des expériences de perte de fonction à l’aide de morpholinos antisens bloquant la traduction de CBFβ ont été réalisées. Ces expériences suggèrent que le facteur CBFβ est nécessaire à la mise en place de la CN et à la différenciation des neurones de Rohon-Beard.
Doctorat en Sciences
info:eu-repo/semantics/nonPublished

Tese de Doutorado apresentada ao Programa de Pós-Graduação em Ciências da Saúde da Universidade do Extremo Sul Catarinense, UNESC, para obtenção do título de Doutora em Ciências da Saúde.
A doença da urina do xarope do bordo (DXB) é um distúrbio neurometabólico de herança autossômica recessiva, causado por uma deficiência da atividade do complexo desidrogenase dos α-cetoácidos de cadeia ramificada. Esta deficiência leva ao acúmulo dos aminoácidos de cadeia ramificada (AACR) leucina, isoleucina e valina, bem como de seus α-cetoácidos correspondentes em tecidos e líquidos corporais de pacientes. O acúmulo destes interfere no metabolismo astrocítico e neuronal, sendo que concentrações elevadas de leucina e do ácido α-cetoisocaproico (CIC) são consideradas particularmente tóxicas ao cérebro. Acredita-se que o CIC seja o mais tóxico dos cetoácidos, visto que esse inibiu o consumo de oxigênio cerebral, induziu estresse oxidativo, provocou deficiência na formação de mielina em cerebelo de ratos e está associado ao aparecimento de sintomas neurológicos. No entanto, até o momento, os mecanismos fisiopatológicos não estão completamente estabelecidos. Considerando que a viabilidade neuronal pode ser afetada pela redução dos fatores neurotróficos e pelo estresse oxidativo, o propósito deste estudo foi avaliar os efeitos neuroquímicos da administração intracerebroventricular (ICV) de CIC sobre estruturas cerebrais de ratos com 30 dias de vida. Analisaram-se os níveis proteicos do fator neurotrófico derivado do cérebro (BDNF), do pró-BDNF e do fator de crescimento neural (NGF). Também foram avaliados os níveis de substâncias reativas ao ácido tiobarbitúrico (TBA-RS), proteínas carboniladas, atividade das enzimas superóxido dismutase (SOD) e catalase (CAT), bem como o dano ao DNA em hipocampo, estriado e córtex cerebral uma hora após a administração ICV de CIC. Os resultados demonstraram que a administração de CIC reduziu os níveis proteicos do BDNF em hipocampo, estriado e córtex cerebral, sem alterar os níveis proteicos de pró-BDNF. Adicionalmente os níveis proteicos de NGF mostraram-se reduzidos em hipocampo, observou-se aumento significativo da concentração do marcador de peroxidação lipídica TBA-RS, bem como da quantidade de proteínas carboniladas em todas as estruturas cerebrais estudadas. A enzima CAT teve sua atividade reduzida no estriado, enquanto que a atividade da SOD se mostrou aumentada em hipocampo e estriado dos animais que receberam CIC. Por fim, a administração ICV de CIC ocasionou aumento significativo do índice e da frequência de danos ao DNA em todas as estruturas estudadas. Em conclusão, esses resultados sugerem que o CIC causa um desequilíbrio nos níveis de neurotrofinas, bem como induz estresse oxidativo. Baseando-se em dados da literatura que demonstram que os metabólitos acumulados na DXB causam desmielinização e prejuízos na memória, especula-se que os efeitos do CIC encontrados neste trabalho possam colaborar para tais achados por causar redução do suporte trófico de BDNF e NGF, pelo estresse oxidativo e dano ao DNA. Além disso, os baixos níveis de BDNF e NGF são consistentes com a hipótese que um déficit nestes fatores neurotróficos pode contribuir para alterações estruturais e funcionais do cérebro subjacentes à fisiopatologia da DXB, apoiando a hipótese do processo neurodegenerativo na DXB.

Made available in DSpace on 2014-06-12T16:00:25Z (GMT). No. of bitstreams: 2 arquivo6557_1.pdf: 4867823 bytes, checksum: 64578a5cde42f460f0745045ec1bb555 (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2007
Conselho Nacional de Desenvolvimento Científico e Tecnológico
Neste trabalho apresentamos um novo modelo neural para segmentação de imagens, baseado nos Mapas Auto-organizáveis SOM (Mapa Auto-organizável - Self-organizing Map) e GWR (Crescer Quando Requerido - Grow When Required) chamado de LARFSOM (Mapa Auto-organizável com Campo Receptivo Adaptativo Local - Local Adaptive Receptive Field Self-organizing Map). As características principais do modelo são: número adaptativo de nodos, topologia variável, inserção de novos nodos baseada em uma medida de similaridade dos protótipos existentes em relação ao padrão de entrada aferida por meio de campo receptivo, remoção de nodos com informações não significativas ao final do treinamento, rápida convergência e baixo custo de processamento para o treinamento. A rede LARFSOM é capaz de segmentar imagens por cor ou por borda: a primeira, é feita através do agrupamento de informações ocorrido no treinamento da rede LAFRSOM seguido de um processo de quantização de cores; já a segunda, ocorre pelo acréscimo de dois nodos RBF (Função de Base Radial - Radial Basis Function) à rede LARFSOM, criando um modelo de dois estágios chamado LARFSOM-RBF. Adicionalmente, o modelo é capaz de salvar em um formato variante do BMP indexado tanto a rede treinada como as informações espaciais dos pixels da imagem. Acrescido de compactação tipo ZIP o arquivo a ser salvo torna-se bem reduzido. Comparações com outros modelos neurais como o SOM, FS-SOM (Mapa Auto-organizável Sensível à Freqüência - Frequency Sensitive Self-organizing Map) e GNG (Gás Neural Crescente - Growing Neural Gas) são feitas mediante segmentação de imagens do mundo real com diferentes níveis de complexidade. Técnicas de processamento de imagens e o formato JPEG são usados para fins de comparação. Os resultados mostram que a rede LARFSOM atinge maior variação de cores da paleta e melhor distribuição espacial 3D RGB das cores selecionadas que os demais modelos. A qualidade das imagens geradas também figura entre os melhores resultados obtidos

Student Number : 9607275H - MSc dissertation - School of Electrical and Information Engineering - Faculty of Engineering and the Built Environment
Internet worms are capable of quickly propagating by exploiting vulnerabilities of hosts that have access to the Internet. Once a computer has been infected, the worms have access to sensitive information on the computer, and are able to corrupt or retransmit this information. This dissertation describes a method of predicting Internet instability due to the presence of a worm on the Internet, using data currently available from global Internet routers. The work is based on previous research which has indicated a link between the increase in the number of Border Gateway Protocol (BGP) routing messages and global Internet instability. The type of system used to provide the prediction is known as an autoencoder. This is a specialised type of neural network, which is able to provide a degree of novelty for inputs. The autoencoder is trained to recognise “normal” data, and therefore provides a high novelty output for inputs dissimilar to the normal data. The BGP Update routing messages sent between routers were used as the only inputs to the autoencoder. These intra-router messages provide route availability information, and inform neighbouring routers of any route changes. The outputs from the network were shown to help provide an early warning mechanism for the presence of a worm. An alternative method for detecting instability is a rule-based system, which generates alarms if the number of certain BGP routing messages exceeds a prespecified threshold. This project compared the autoencoder to a simple rule-based system. The results showed that the autoencoder provided a better prediction and was less complex for a network administrator to configure. Although the correlation between the number of BGP Updates and global Internet instability has been shown previously, this work presents the first known application of a neural network to predict the instability using this correlation. A system based on this strategy has the potential to reduce the damage done by a worm’s propagation and payload, by providing an automated means of detection that is faster than that of a human.

“Winner-take-all” networks typically pick as winners that alternative with the largest excitatory input. This choice is far from optimal when there is uncertainty in the strength of the inputs, and when information is available about how alternatives may be related. In the Social Choice community, many other procedures will yield more robust winners. The Borda Count and the pair-wise Condorcet tally are among the most favored. Their implementations are simple modifications of classical recurrent networks.

Dissertação de Mestrado Integrado em Engenharia Electrotécnica e de Computadores apresentada à Faculdade de Ciências e Tecnologia
Os dados gerados por anualmente rondam os 40 trilhões de gigabytes. Este aumento significativo de dados todos os anos trás a necessidade de assegurar a proteção de informação sensível. A Inteligência Artificial tem vindo a melhorar cada vez mais os seus resultados, apresentando modelos capazes de responder rigorosamente em áreas de atuação críticas, por exemplo, medicina, veículos autónomos, robótica, etc. Estes algoritmos precisam de enormes quantidades de dados disponíveis para otimizarem ao máximo a sua resposta perante todos a sua área de operação.Surgiu a necessidade de continuar a melhorar estes algoritmos mantendo a privacidade e confidencialidade dos dados utilizados.Desta forma, foi criado o conceito de Federated Learning. O Federated Learning permite continuar a treinar algoritmos de Machine Learning sem partilhar os dados utilizados para a convergência do modelo. O Federated Learning apresenta apresenta algumas similaridades com o Distributed Learning. Em ambos os conceitos o treino é distribuido, no entanto o Federated Learning descentraliza também os dados de forma a manter a informação privada.O objetivo desta dissertação passa por explorar o conceito de Federated Learning, assim como comparar diretamente este conceito com o Machine Learning centralizado. Para tal, é mostrada a arquitetura necessária para a construção de uma solução federada. Este documento apresenta ainda resultados obtidos com soluções federadas tanto em ambiente de simulação como numa implementação em ambiente real. Finalmente, é também apresentado um ponto de vista dos resultados obtidos e opções de otimização de uma solução com Federated Learning são discutidas.
The data generated annually is around 40 trillion gigabytes. This significant increase in data every year brings with it the need to ensure the protection of sensitive information. Artificial Intelligence has been improving its results more and more, presenting models capable of responding rigorously in critical areas, for example medicine, autonomous vehicles, robotics, etc. These algorithms need huge amounts of available data to optimize their response to all their area of operation.The urge to continue to improve these algorithms while maintaining the privacy and confidentiality of the data used emerged.Thus, the concept of Federated Learning was created. Federated Learning allows to continue training Machine Learning algorithms without sharing the data used for model convergence. Federated Learning has some similarities with Distributed Learning. In both concepts the training is distributed, however, Federated Learning also decentralizes the data in order to keep the information private.The objective of this dissertation is to explore the concept of Federated Learning, as well as to directly compare this concept with centralized Machine Learning. To this end, the architecture required to build a federated solution is analyzed in depth. This dissertation also presents results obtained with federated solutions in both simulation and real-world deployment. Finally, a viewpoint of the obtained results is also presented, and options for optimizing a solution with Federated Learning are discussed.

Nous étudions dans cette thèse comment les réseaux de neurones profonds peuvent être utilisés dans différents domaines de la vision artificielle. La vision artificielle est un domaine interdisciplinaire qui traite de la compréhension d’images et de vidéos numériques. Les problèmes de ce domaine ont traditionnellement été adressés avec des méthodes ad-hoc nécessitant beaucoup de réglages manuels. En effet, ces systèmes de vision artificiels comprenaient jusqu’à récemment une série de modules optimisés indépendamment. Cette approche est très raisonnable dans la mesure où, avec peu de données, elle bénéficient autant que possible des connaissances du chercheur. Mais cette avantage peut se révéler être une limitation si certaines données d’entré n’ont pas été considérées dans la conception de l’algorithme. Avec des volumes et une diversité de données toujours plus grands, ainsi que des capacités de calcul plus rapides et économiques, les réseaux de neurones profonds optimisés d’un bout à l’autre sont devenus une alternative attrayante. Nous démontrons leur avantage avec une série d’articles de recherche, chacun d’entre eux trouvant une solution à base de réseaux de neurones profonds à un problème d’analyse ou de synthèse visuelle particulier. Dans le premier article, nous considérons un problème de vision classique: la détection de bords et de contours. Nous partons de l’approche classique et la rendons plus ‘neurale’ en combinant deux étapes, la détection et la description de motifs visuels, en un seul réseau convolutionnel. Cette méthode, qui peut ainsi s’adapter à de nouveaux ensembles de données, s’avère être au moins aussi précis que les méthodes conventionnelles quand il s’agit de domaines qui leur sont favorables, tout en étant beaucoup plus robuste dans des domaines plus générales. Dans le deuxième article, nous construisons une nouvelle architecture pour la manipulation d’images qui utilise l’idée que la majorité des pixels produits peuvent d’être copiés de l’image d’entrée. Cette technique bénéficie de plusieurs avantages majeurs par rapport à l’approche conventionnelle en apprentissage profond. En effet, elle conserve les détails de l’image d’origine, n’introduit pas d’aberrations grâce à la capacité limitée du réseau sous-jacent et simplifie l’apprentissage. Nous démontrons l’efficacité de cette architecture dans le cadre d’une tâche de correction du regard, où notre système produit d’excellents résultats. Dans le troisième article, nous nous éclipsons de la vision artificielle pour étudier le problème plus générale de l’adaptation à de nouveaux domaines. Nous développons un nouvel algorithme d’apprentissage, qui assure l’adaptation avec un objectif auxiliaire à la tâche principale. Nous cherchons ainsi à extraire des motifs qui permettent d’accomplir la tâche mais qui ne permettent pas à un réseau dédié de reconnaître le domaine. Ce réseau est optimisé de manière simultané avec les motifs en question, et a pour tâche de reconnaître le domaine de provenance des motifs. Cette technique est simple à implémenter, et conduit pourtant à l’état de l’art sur toutes les tâches de référence. Enfin, le quatrième article présente un nouveau type de modèle génératif d’images. À l’opposé des approches conventionnels à base de réseaux de neurones convolutionnels, notre système baptisé SPIRAL décrit les images en termes de programmes bas-niveau qui sont exécutés par un logiciel de graphisme ordinaire. Entre autres, ceci permet à l’algorithme de ne pas s’attarder sur les détails de l’image, et de se concentrer plutôt sur sa structure globale. L’espace latent de notre modèle est, par construction, interprétable et permet de manipuler des images de façon prévisible. Nous montrons la capacité et l’agilité de cette approche sur plusieurs bases de données de référence.
In the present thesis, we study how deep neural networks can be applied to various tasks in computer vision. Computer vision is an interdisciplinary field that deals with understanding of digital images and video. Traditionally, the problems arising in this domain were tackled using heavily hand-engineered adhoc methods. A typical computer vision system up until recently consisted of a sequence of independent modules which barely talked to each other. Such an approach is quite reasonable in the case of limited data as it takes major advantage of the researcher's domain expertise. This strength turns into a weakness if some of the input scenarios are overlooked in the algorithm design process. With the rapidly increasing volumes and varieties of data and the advent of cheaper and faster computational resources end-to-end deep neural networks have become an appealing alternative to the traditional computer vision pipelines. We demonstrate this in a series of research articles, each of which considers a particular task of either image analysis or synthesis and presenting a solution based on a ``deep'' backbone. In the first article, we deal with a classic low-level vision problem of edge detection. Inspired by a top-performing non-neural approach, we take a step towards building an end-to-end system by combining feature extraction and description in a single convolutional network. The resulting fully data-driven method matches or surpasses the detection quality of the existing conventional approaches in the settings for which they were designed while being significantly more usable in the out-of-domain situations. In our second article, we introduce a custom architecture for image manipulation based on the idea that most of the pixels in the output image can be directly copied from the input. This technique bears several significant advantages over the naive black-box neural approach. It retains the level of detail of the original images, does not introduce artifacts due to insufficient capacity of the underlying neural network and simplifies training process, to name a few. We demonstrate the efficiency of the proposed architecture on the challenging gaze correction task where our system achieves excellent results. In the third article, we slightly diverge from pure computer vision and study a more general problem of domain adaption. There, we introduce a novel training-time algorithm (\ie, adaptation is attained by using an auxilliary objective in addition to the main one). We seek to extract features that maximally confuse a dedicated network called domain classifier while being useful for the task at hand. The domain classifier is learned simultaneosly with the features and attempts to tell whether those features are coming from the source or the target domain. The proposed technique is easy to implement, yet results in superior performance in all the standard benchmarks. Finally, the fourth article presents a new kind of generative model for image data. Unlike conventional neural network based approaches our system dubbed SPIRAL describes images in terms of concise low-level programs executed by off-the-shelf rendering software used by humans to create visual content. Among other things, this allows SPIRAL not to waste its capacity on minutae of datasets and focus more on the global structure. The latent space of our model is easily interpretable by design and provides means for predictable image manipulation. We test our approach on several popular datasets and demonstrate its power and flexibility.