Advances in unsupervised learning and deep neural networks have led to increased performance in a number of domains, and to the ability to draw strong comparisons between the biological method of self-organization conducted by the brain and computational mechanisms. This thesis aims to use real-world data to tackle two areas in the domain of computer vision which have biological equivalents: feature detection and motion tracking. The aforementioned advances have allowed efficient learning of feature representations directly from large sets of unlabeled data instead of using traditional handcrafted features. The first part of this thesis evaluates such representations by comparing regularization and preprocessing methods which incorporate local neighbouring information during training on a single-layer neural network. The networks are trained and tested on the Hollywood2 video dataset, as well as the static CIFAR-10, STL-10, COIL-100, and MNIST image datasets. The induction of topography or simple image blurring via Gaussian filters during training produces better discriminative features as evidenced by the consistent and notable increase in classification results that they produce. In the visual domain, invariant features are desirable such that objects can be classified despite transformations. It is found that most of the compared methods produce more invariant features, however, classification accuracy does not correlate to invariance. The second, and paramount, contribution of this thesis is a biologically-inspired model to explain the emergence of motion tracking behaviour in early development using unsupervised learning. The model’s self-organization is biased by an original concept called retinal constancy, which measures how similar visual contents are between successive frames. In the proposed two-layer deep network, when exposed to real-world video, the first layer learns to encode visual motion, and the second layer learns to relate that motion to gaze movements, which it perceives and creates through bi-directional nodes. This is unique because it uses general machine learning algorithms, and their inherent generative properties, to learn from real-world data. It also implements a biological theory and learns in a fully unsupervised manner. An analysis of its parameters and limitations is conducted, and its tracking performance is evaluated. Results show that this model is able to successfully follow targets in real-world video, despite being trained without supervision on real-world video.

Machine learning is evidenced as a research area with the main purpose of developing computational methods that are capable of learning with their previously acquired experiences. Although a large amount of machine learning techniques has been proposed and successfully applied in real systems, there are still many challenging issues, which need be addressed. In the last years, an increasing interest in techniques based on complex networks (large-scale graphs with nontrivial connection patterns) has been verified. This emergence is explained by the inherent advantages provided by the complex network representation, which is able to capture the spatial, topological and functional relations of the data. In this work, we investigate the new features and possible advantages offered by complex networks in the machine learning domain. In fact, we do show that the network-based approach really brings interesting features for supervised, semisupervised, and unsupervised learning. Specifically, we reformulate a previously proposed particle competition technique for both unsupervised and semisupervised learning using a stochastic nonlinear dynamical system. Moreover, an analytical analysis is supplied, which enables one to predict the behavior of the proposed technique. In addition to that, data reliability issues are explored in semisupervised learning. Such matter has practical importance and is found to be of little investigation in the literature. With the goal of validating these techniques for solving real problems, simulations on broadly accepted databases are conducted. Still in this work, we propose a hybrid supervised classification technique that combines both low and high orders of learning. The low level term can be implemented by any classification technique, while the high level term is realized by the extraction of features of the underlying network constructed from the input data. Thus, the former classifies the test instances by their physical features, while the latter measures the compliance of the test instances with the pattern formation of the data. Our study shows that the proposed technique not only can realize classification according to the semantic meaning of the data, but also is able to improve the performance of traditional classification techniques. Finally, it is expected that this study will contribute, in a relevant manner, to the machine learning area / Aprendizado de máquina figura-se como uma área de pesquisa que visa a desenvolver métodos computacionais capazes de aprender com a experiência. Embora uma grande quantidade de técnicas de aprendizado de máquina foi proposta e aplicada, com sucesso, em sistemas reais, existem ainda inúmeros problemas desafiantes que necessitam ser explorados. Nos últimos anos, um crescente interesse em técnicas baseadas em redes complexas (grafos de larga escala com padrões de conexão não triviais) foi verificado. Essa emergência é explicada pelas inerentes vantagens que a representação em redes complexas traz, sendo capazes de capturar as relações espaciais, topológicas e funcionais dos dados. Nesta tese, serão investigadas as possíveis vantagens oferecidas por redes complexas quando utilizadas no domínio de aprendizado de máquina. De fato, será mostrado que a abordagem por redes realmente proporciona melhorias nos aprendizados supervisionado, semissupervisionado e não supervisionado. Especificamente, será reformulada uma técnica de competição de partículas para o aprendizado não supervisionado e semissupervisionado por meio da utilização de um sistema dinâmico estocástico não linear. Em complemento, uma análise analítica de tal modelo será desenvolvida, permitindo o entendimento evolucional do modelo no tempo. Além disso, a questão de confiabilidade de dados será investigada no aprendizado semissupervisionado. Tal tópico tem importância prática e é pouco estudado na literatura. Com o objetivo de validar essas técnicas em problemas reais, simulações computacionais em bases de dados consagradas pela literatura serão conduzidas. Ainda nesse trabalho, será proposta uma técnica híbrica de classificação supervisionada que combina tanto o aprendizado de baixo como de alto nível. O termo de baixo nível pode ser implementado por qualquer técnica de classificação tradicional, enquanto que o termo de alto nível é realizado pela extração das características de uma rede construída a partir dos dados de entrada. Nesse contexto, aquele classifica as instâncias de teste segundo qualidades físicas, enquanto que esse estima a conformidade da instância de teste com a formação de padrões dos dados. Os estudos aqui desenvolvidos mostram que o método proposto pode melhorar o desempenho de técnicas tradicionais de classificação, além de permitir uma classificação de acordo com o significado semântico dos dados. Enfim, acredita-se que este estudo possa gerar contribuições relevantes para a área de aprendizado de máquina.

Humans' incredible capacity to model the complexity of the physical world is possible because they cast this complexity as the composition of simpler entities and rules to process them. Extensive work in cognitive science indeed shows that human perception and reasoning ability is structured around objects. Motivated by this observation, a growing number of recent work focused on entity-centric approaches to learning representation and their potential to facilitate downstream tasks. In the first contribution, we show how an entity-centric approach to learning a transition model allows us to extract meaningful visual entities and to learn transition rules that achieve better compositional generalization. In the second contribution, we show how an entity-centric approach to generating graphs allows us to design a model for conditional graph generation that permits direct optimisation of the graph properties. We investigate the performance of our model in a prototype-based molecular graph generation task. In this task, called lead optimization in drug discovery, we wish to adjust a few physico-chemical properties of a molecule that has proven efficient in vitro in order to make a drug out of it. / L'incroyable capacité des humains à modéliser la complexité du monde physique est rendue possible par la décomposition qu'ils en font en un ensemble d'entités et de règles simples. De nombreux travaux en sciences cognitives montre que la perception humaine et sa capacité à raisonner est essentiellement centrée sur la notion d'objet. Motivés par cette observation, de récents travaux se sont intéressés aux différentes approches d'apprentissage de représentations centrées sur des entités et comment ces représentations peuvent être utilisées pour résoudre plus facilement des tâches sous-jacentes. Dans la première contribution on montre comment une architecture centrée sur la notion d'entité va permettre d'extraire des entités visuelles interpretables et d'apprendre un modèle du monde plus robuste aux différentes configurations d'objets. Dans la deuxième contribution on s’intéresse à un modèle de génération de graphes dont l'architecture est également centrée sur la notion d'entités et comment cette architecture rend plus facile l'apprentissage d'une génération conditionelle à certaines propriétés du graphe. On s’intéresse plus particulièrement aux applications en découverte de médicaments. Dans cette tâche, on souhaite optimiser certaines propriétés physico-chmiques du graphe d'une molécule qui a été efficace in-vitro et dont on veut faire un médicament.

No description available.

No description available.

No description available.

No description available.

<div>With more and more videos taken from dash cams on thousands of cars, retrieving these videos and searching for important information is a daunting task. The purpose of this work is to mine some key road and vehicle motion attributes in a large-scale driving video data set for traffic analysis, sensing algorithm development and autonomous driving test benchmarks. Current sensing and control of autonomous cars based on full-view identification makes it difficult to maintain a high-frequency with a fast-moving vehicle, since computation is increasingly used to cope with driving environment changes.</div><div><br></div><div>A big challenge in video data mining is how to deal with huge amounts of data. We use a compact representation called the road profile system to visualize the road environment in long 2D images. It reduces the data from each frame of image to one line, thereby compressing the video clip to the image. This data dimensionality reduction method has several advantages: First, the data size is greatly compressed. The data is compressed from a video to an image, and each frame in the video is compressed into a line. The data size is compressed hundreds of times. While the size and dimensionality of the data has been compressed greatly, the useful information in the driving video is still completely preserved, and motion information is even better represented more intuitively. Because of the data and dimensionality reduction, the identification algorithm computational efficiency is higher than the full-view identification method, and it makes the real-time identification on road is possible. Second, the data is easier to be visualized, because the data is reduced in dimensionality, and the three-dimensional video data is compressed into two-dimensional data, the reduction is more conducive to the visualization and mutual comparison of the data. Third, continuously changing attributes are easier to show and be captured. Due to the more convenient visualization of two-dimensional data, the position, color and size of the same object within a few frames will be easier to compare and capture. At the same time, in many cases, the trouble caused by tracking and matching can be eliminated. Based on the road profile system, there are three tasks in autonomous driving are achieved using the road profile images.</div><div><br></div><div>The first application is road edge detection under different weather and appearance for road following in autonomous driving to capture the road profile image and linearity profile image in the road profile system. This work uses naturalistic driving video data mining to study the appearance of roads, which covers large-scale road data and changes. This work excavated a large number of naturalistic driving video sets to sample the light-sensitive area for color feature distribution. The effective road contour image is extracted from the long-time driving video, thereby greatly reducing the amount of video data. Then, the weather and lighting type can be identified. For each weather and lighting condition obvious features are I identified at the edge of the road to distinguish the road edge. </div><div><br></div><div>The second application is detecting vehicle interactions in driving videos via motion profile images to capture the motion profile image in the road profile system. This work uses visual actions recorded in driving videos taken by a dashboard camera to identify this interaction. The motion profile images of the video are filtered at key locations, thereby reducing the complexity of object detection, depth sensing, target tracking and motion estimation. The purpose of this reduction is for decision making of vehicle actions such as lane changing, vehicle following, and cut-in handling.</div><div><br></div><div>The third application is motion planning based on vehicle interactions and driving video. Taking note of the fact that a car travels in a straight line, we simply identify a few sample lines in the view to constantly scan the road, vehicles, and environment, generating a portion of the entire video data. Without using redundant data processing, we performed semantic segmentation to streaming road profile images. We plan the vehicle's path/motion using the smallest data set possible that contains all necessary information for driving.</div><div><br></div><div>The results are obtained efficiently, and the accuracy is acceptable. The results can be used for driving video mining, traffic analysis, driver behavior understanding, etc.</div>

Un objectif de longue date dans le domaine de la vision par ordinateur est de déduire le contenu 3D d’une scène à partir d’une seule photo, une tâche connue sous le nom d’inverse graphics. L’apprentissage automatique a, dans les dernières années, permis à de nombreuses approches de faire de grands progrès vers la résolution de ce problème. Cependant, la plupart de ces approches requièrent des données de supervision 3D qui sont coûteuses et parfois impossible à obtenir, ce qui limite les capacités d’apprentissage de telles œuvres. Dans ce travail, nous explorons l’architecture des méthodes d’inverse graphics non-supervisées et proposons deux méthodes basées sur des représentations 3D et algorithmes de rendus différentiables distincts: les surfels ainsi qu’une nouvelle représentation basée sur Voronoï. Dans la première méthode basée sur les surfels, nous montrons que, bien qu’efficace pour maintenir la cohérence visuelle, la production de surfels à l’aide d’une carte de profondeur apprise entraîne des ambiguïtés car la relation entre la carte de profondeur et le rendu n’est pas bijective. Dans notre deuxième méthode, nous introduisons une nouvelle représentation 3D basée sur les diagrammes de Voronoï qui modélise des objets/scènes à la fois explicitement et implicitement, combinant ainsi les avantages des deux approches. Nous montrons comment cette représentation peut être utilisée à la fois dans un contexte supervisé et non-supervisé et discutons de ses avantages par rapport aux représentations 3D traditionnelles / A long standing goal of computer vision is to infer the underlying 3D content in a scene from a single photograph, a task known as inverse graphics. Machine learning has, in recent years, enabled many approaches to make great progress towards solving this problem. However, most approaches rely on 3D supervision data which is expensive and sometimes impossible to obtain and therefore limits the learning capabilities of such work. In this work, we explore the deep unsupervised inverse graphics training pipeline and propose two methods based on distinct 3D representations and associated differentiable rendering algorithms: namely surfels and a novel Voronoi-based representation. In the first method based on surfels, we show that, while effective at maintaining view-consistency, producing view-dependent surfels using a learned depth map results in ambiguities as the mapping between depth map and rendering is non-bijective. In our second method, we introduce a novel 3D representation based on Voronoi diagrams which models objects/scenes both explicitly and implicitly simultaneously, thereby combining the benefits of both. We show how this representation can be used in both a supervised and unsupervised context and discuss its advantages compared to traditional 3D representations.

Les modèles de réseaux de neurones et en particulier les modèles profonds sont aujourd'hui l'un des modèles à l'état de l'art en apprentissage automatique et ses applications. Les réseaux de neurones profonds récents possèdent de nombreuses couches cachées ce qui augmente significativement le nombre total de paramètres. L'apprentissage de ce genre de modèles nécessite donc un grand nombre d'exemples étiquetés, qui ne sont pas toujours disponibles en pratique. Le sur-apprentissage est un des problèmes fondamentaux des réseaux de neurones, qui se produit lorsque le modèle apprend par coeur les données d'apprentissage, menant à des difficultés à généraliser sur de nouvelles données. Le problème du sur-apprentissage des réseaux de neurones est le thème principal abordé dans cette thèse. Dans la littérature, plusieurs solutions ont été proposées pour remédier à ce problème, tels que l'augmentation de données, l'arrêt prématuré de l'apprentissage ("early stopping"), ou encore des techniques plus spécifiques aux réseaux de neurones comme le "dropout" ou la "batch normalization". Dans cette thèse, nous abordons le sur-apprentissage des réseaux de neurones profonds sous l'angle de l'apprentissage de représentations, en considérant l'apprentissage avec peu de données. Pour aboutir à cet objectif, nous avons proposé trois différentes contributions. La première contribution, présentée dans le chapitre 2, concerne les problèmes à sorties structurées dans lesquels les variables de sortie sont à grande dimension et sont généralement liées par des relations structurelles. Notre proposition vise à exploiter ces relations structurelles en les apprenant de manière non-supervisée avec des autoencodeurs. Nous avons validé notre approche sur un problème de régression multiple appliquée à la détection de points d'intérêt dans des images de visages. Notre approche a montré une accélération de l'apprentissage des réseaux et une amélioration de leur généralisation. La deuxième contribution, présentée dans le chapitre 3, exploite la connaissance a priori sur les représentations à l'intérieur des couches cachées dans le cadre d'une tâche de classification. Cet à priori est basé sur la simple idée que les exemples d'une même classe doivent avoir la même représentation interne. Nous avons formalisé cet à priori sous la forme d'une pénalité que nous avons rajoutée à la fonction de perte. Des expérimentations empiriques sur la base MNIST et ses variantes ont montré des améliorations dans la généralisation des réseaux de neurones, particulièrement dans le cas où peu de données d'apprentissage sont utilisées. Notre troisième et dernière contribution, présentée dans le chapitre 4, montre l'intérêt du transfert d'apprentissage ("transfer learning") dans des applications dans lesquelles peu de données d'apprentissage sont disponibles. L'idée principale consiste à pré-apprendre les filtres d'un réseau à convolution sur une tâche source avec une grande base de données (ImageNet par exemple), pour les insérer par la suite dans un nouveau réseau sur la tâche cible. Dans le cadre d'une collaboration avec le centre de lutte contre le cancer "Henri Becquerel de Rouen", nous avons construit un système automatique basé sur ce type de transfert d'apprentissage pour une application médicale où l'on dispose d’un faible jeu de données étiquetées. Dans cette application, la tâche consiste à localiser la troisième vertèbre lombaire dans un examen de type scanner. L’utilisation du transfert d’apprentissage ainsi que de prétraitements et de post traitements adaptés a permis d’obtenir des bons résultats, autorisant la mise en oeuvre du modèle en routine clinique. / Neural network models and deep models are one of the leading and state of the art models in machine learning. They have been applied in many different domains. Most successful deep neural models are the ones with many layers which highly increases their number of parameters. Training such models requires a large number of training samples which is not always available. One of the fundamental issues in neural networks is overfitting which is the issue tackled in this thesis. Such problem often occurs when the training of large models is performed using few training samples. Many approaches have been proposed to prevent the network from overfitting and improve its generalization performance such as data augmentation, early stopping, parameters sharing, unsupervised learning, dropout, batch normalization, etc. In this thesis, we tackle the neural network overfitting issue from a representation learning perspective by considering the situation where few training samples are available which is the case of many real world applications. We propose three contributions. The first one presented in chapter 2 is dedicated to dealing with structured output problems to perform multivariate regression when the output variable y contains structural dependencies between its components. Our proposal aims mainly at exploiting these dependencies by learning them in an unsupervised way. Validated on a facial landmark detection problem, learning the structure of the output data has shown to improve the network generalization and speedup its training. The second contribution described in chapter 3 deals with the classification task where we propose to exploit prior knowledge about the internal representation of the hidden layers in neural networks. This prior is based on the idea that samples within the same class should have the same internal representation. We formulate this prior as a penalty that we add to the training cost to be minimized. Empirical experiments over MNIST and its variants showed an improvement of the network generalization when using only few training samples. Our last contribution presented in chapter 4 showed the interest of transfer learning in applications where only few samples are available. The idea consists in re-using the filters of pre-trained convolutional networks that have been trained on large datasets such as ImageNet. Such pre-trained filters are plugged into a new convolutional network with new dense layers. Then, the whole network is trained over a new task. In this contribution, we provide an automatic system based on such learning scheme with an application to medical domain. In this application, the task consists in localizing the third lumbar vertebra in a 3D CT scan. A pre-processing of the 3D CT scan to obtain a 2D representation and a post-processing to refine the decision are included in the proposed system. This work has been done in collaboration with the clinic "Rouen Henri Becquerel Center" who provided us with data