Diese Arbeit untersucht Inferenz für Streuungsparameter bedingter Gaußprozesse anhand diskreter verrauschter Beobachtungen in einem Hochfrequenz-Setting. Unser Ziel dabei ist es, eine asymptotische Charakterisierung von effizienter Schätzung in einem allgemeine Gaußschen Rahmen zu finden. Für ein parametrisches Fundamentalmodell wird ein Hájek-Le Cam-Faltungssatz hergeleitet, welcher eine exakte asymptotische untere Schranke für Schätzmethoden liefert. Dazu passende obere Schranken werden konstruiert und die Bedeutung des Satzes wird verdeutlicht anhand zahlreicher Beispiele wie der (fraktionellen) Brownschen Bewegung, dem Ornstein-Uhlenbeck-Prozess oder integrierten Prozessen. Die Herleitung der Effizienzresultate basiert auf asymptotischen Äquivalenzen und kann für verschiedene Verallgemeinerungen des parametrischen Fundamentalmodells verwendet werden. Als eine solche Erweiterung betrachten wir das Schätzen der quadrierten Kovariation eines stetigen Martingals anhand verrauschter asynchroner Beobachtungen, welches ein fundamentales Schätzproblem in der Öknometrie ist. Für dieses Modell erhalten wir einen semi-parametrischen Faltungssatz, welcher bisherige Resultate im Sinne von Multidimensionalität, Asynchronität und Annahmen verallgemeinert. Basierend auf den vorhergehenden Herleitungen entwickeln wir einen statistischen Test für den Hurst-Parameter einer fraktionellen Brownschen Bewegung. Ein Score- und ein Likelihood-Quotienten-Test werden implementiert sowie analysiert und erste empirische Eindrücke vermittelt. / This work studies inference on scaling parameters of a conditionally Gaussian process under discrete noisy observations in a high-frequency regime. Our aim is to find an asymptotic characterisation of efficient estimation for a general Gaussian framework. For a parametric basic case model a Hájek-Le Cam convolution theorem is derived, yielding an exact asymptotic lower bound for estimators. Matching upper bounds are constructed and the importance of the theorem is illustrated by various examples of interest such as the (fractional) Brownian motion, the Ornstein-Uhlenbeck process or integrated processes. The derivation of the efficiency result is based on asymptotic equivalences and can be employed for several generalisations of the parametric basic case model. As such an extension we consider estimation of the quadratic covariation of a continuous martingale from noisy asynchronous observations, which is a fundamental estimation problem in econometrics. For this model, a semi-parametric convolution theorem is obtained which generalises existing results in terms of multidimensionality, asynchronicity and assumptions. Based on the previous derivations, we develop statistical tests on the Hurst parameter of a fractional Brownian motion. A score test and a likelihood ratio type test are implemented as well as analysed and first empirical impressions are given.

A statistical meta-analysis combines the results of several studies that address a set of related research hypotheses, thus increasing the power and reliability of the inference. Meta-analytic methods are over 50 years old and play an important role in science; pooling evidence from many trials to provide answers that any one trial would have insufficient samples to address. On the other hand, the number of neuroimaging studies is growing dramatically, with many of these publications containing conflicting results, or being based on only a small number of subjects. Hence there has been increasing interest in using meta-analysis methods to find consistent results for a specific functional task, or for predicting the results of a study that has not been performed directly. Current state of neuroimaging meta-analysis is limited to coordinate-based meta-analysis (CBMA), i.e., using only the coordinates of activation peaks that are reported by a group of studies, in order to "localize" the brain regions that respond to a certain type of stimulus. This class of meta-analysis suffers from a series of problems and hence cannot result in as accurate results as desired. In this research, we describe the problems that existing CBMA methods are suffering from and introduce a hierarchical mixed-effects image-based metaanalysis (IBMA) solution that incorporates the sufficient statistics (i.e., voxel-wise effect size and its associated uncertainty) from each study. In order to improve the statistical-inference stage of our proposed IBMA method, we introduce a nonparametric technique that is capable of adjusting such an inference for spatial nonstationarity. Given that in common practice, neuroimaging studies rarely provide the full image data, in an attempt to improve the existing CBMA techniques we introduce a fully automatic model-based approach that employs Gaussian-process regression (GPR) for estimating the meta-analytic statistic image from its corresponding sparse and noisy observations (i.e., the collected foci). To conclude, we introduce a new way to approach neuroimaging meta-analysis that enables the analysis to result in information such as “functional connectivity” and networks of the brain regions’ interactions, rather than just localizing the functions.

Part I - Level crossings and other level functionals.<br />Part II - Some contributions in statistics of extremes and in statistical mechanics.

Cette thèse envisage un ensemble de méthodes permettant aux algorithmes d'apprentissage statistique de mieux traiter la nature séquentielle des problèmes de gestion de portefeuilles financiers. Nous débutons par une considération du problème général de la composition d'algorithmes d'apprentissage devant gérer des tâches séquentielles, en particulier celui de la mise-à-jour efficace des ensembles d'apprentissage dans un cadre de validation séquentielle. Nous énumérons les desiderata que des primitives de composition doivent satisfaire, et faisons ressortir la difficulté de les atteindre de façon rigoureuse et efficace. Nous poursuivons en présentant un ensemble d'algorithmes qui atteignent ces objectifs et présentons une étude de cas d'un système complexe de prise de décision financière utilisant ces techniques. Nous décrivons ensuite une méthode générale permettant de transformer un problème de décision séquentielle non-Markovien en un problème d'apprentissage supervisé en employant un algorithme de recherche basé sur les K meilleurs chemins. Nous traitons d'une application en gestion de portefeuille où nous entraînons un algorithme d'apprentissage à optimiser directement un ratio de Sharpe (ou autre critère non-additif incorporant une aversion au risque). Nous illustrons l'approche par une étude expérimentale approfondie, proposant une architecture de réseaux de neurones spécialisée à la gestion de portefeuille et la comparant à plusieurs alternatives. Finalement, nous introduisons une représentation fonctionnelle de séries chronologiques permettant à des prévisions d'être effectuées sur un horizon variable, tout en utilisant un ensemble informationnel révélé de manière progressive. L'approche est basée sur l'utilisation des processus Gaussiens, lesquels fournissent une matrice de covariance complète entre tous les points pour lesquels une prévision est demandée. Cette information est utilisée à bon escient par un algorithme qui transige activement des écarts de cours (price spreads) entre des contrats à terme sur commodités. L'approche proposée produit, hors échantillon, un rendement ajusté pour le risque significatif, après frais de transactions, sur un portefeuille de 30 actifs. / This thesis considers a number of approaches to make machine learning algorithms better suited to the sequential nature of financial portfolio management tasks. We start by considering the problem of the general composition of learning algorithms that must handle temporal learning tasks, in particular that of creating and efficiently updating the training sets in a sequential simulation framework. We enumerate the desiderata that composition primitives should satisfy, and underscore the difficulty of rigorously and efficiently reaching them. We follow by introducing a set of algorithms that accomplish the desired objectives, presenting a case-study of a real-world complex learning system for financial decision-making that uses those techniques. We then describe a general method to transform a non-Markovian sequential decision problem into a supervised learning problem using a K-best paths search algorithm. We consider an application in financial portfolio management where we train a learning algorithm to directly optimize a Sharpe Ratio (or other risk-averse non-additive) utility function. We illustrate the approach by demonstrating extensive experimental results using a neural network architecture specialized for portfolio management and compare against well-known alternatives. Finally, we introduce a functional representation of time series which allows forecasts to be performed over an unspecified horizon with progressively-revealed information sets. By virtue of using Gaussian processes, a complete covariance matrix between forecasts at several time-steps is available. This information is put to use in an application to actively trade price spreads between commodity futures contracts. The approach delivers impressive out-of-sample risk-adjusted returns after transaction costs on a portfolio of 30 spreads.

Series data, sequences of measured values, are ubiquitous. Whenever observations are made along a path in space or time, a data sequence results. To comprehend nature and shape it to our will, or to make informed decisions based on what we know, we need methods to make sense of such data. Of particular interest are probabilistic descriptions, which enable us to represent uncertainty and random variation inherent to the world around us. This thesis presents and expands upon some tools for creating probabilistic models of sequences, with an eye towards applications involving speech and language. Modelling speech and language is not only of use for creating listening, reading, talking, and writing machines---for instance allowing human-friendly interfaces to future computational intelligences and smart devices of today---but probabilistic models may also ultimately tell us something about ourselves and the world we occupy. The central theme of the thesis is the creation of new or improved models more appropriate for our intended applications, by weakening limiting and questionable assumptions made by standard modelling techniques. One contribution of this thesis examines causal-state splitting reconstruction (CSSR), an algorithm for learning discrete-valued sequence models whose states are minimal sufficient statistics for prediction. Unlike many traditional techniques, CSSR does not require the number of process states to be specified a priori, but builds a pattern vocabulary from data alone, making it applicable for language acquisition and the identification of stochastic grammars. A paper in the thesis shows that CSSR handles noise and errors expected in natural data poorly, but that the learner can be extended in a simple manner to yield more robust and stable results also in the presence of corruptions. Even when the complexities of language are put aside, challenges remain. The seemingly simple task of accurately describing human speech signals, so that natural synthetic speech can be generated, has proved difficult, as humans are highly attuned to what speech should sound like. Two papers in the thesis therefore study nonparametric techniques suitable for improved acoustic modelling of speech for synthesis applications. Each of the two papers targets a known-incorrect assumption of established methods, based on the hypothesis that nonparametric techniques can better represent and recreate essential characteristics of natural speech. In the first paper of the pair, Gaussian process dynamical models (GPDMs), nonlinear, continuous state-space dynamical models based on Gaussian processes, are shown to better replicate voiced speech, without traditional dynamical features or assumptions that cepstral parameters follow linear autoregressive processes. Additional dimensions of the state-space are able to represent other salient signal aspects such as prosodic variation. The second paper, meanwhile, introduces KDE-HMMs, asymptotically-consistent Markov models for continuous-valued data based on kernel density estimation, that additionally have been extended with a fixed-cardinality discrete hidden state. This construction is shown to provide improved probabilistic descriptions of nonlinear time series, compared to reference models from different paradigms. The hidden state can be used to control process output, making KDE-HMMs compelling as a probabilistic alternative to hybrid speech-synthesis approaches. A final paper of the thesis discusses how models can be improved even when one is restricted to a fundamentally imperfect model class. Minimum entropy rate simplification (MERS), an information-theoretic scheme for postprocessing models for generative applications involving both speech and text, is introduced. MERS reduces the entropy rate of a model while remaining as close as possible to the starting model. This is shown to produce simplified models that concentrate on the most common and characteristic behaviours, and provides a continuum of simplifications between the original model and zero-entropy, completely predictable output. As the tails of fitted distributions may be inflated by noise or empirical variability that a model has failed to capture, MERS's ability to concentrate on high-probability output is also demonstrated to be useful for denoising models trained on disturbed data. / <p>QC 20131128</p> / ACORNS: Acquisition of Communication and Recognition Skills / LISTA – The Listening Talker

Cette thèse est une contribution à la résolution du problème d’optimisation sous contrainte de fiabilité. Cette méthode de dimensionnement probabiliste vise à prendre en compte les incertitudes inhérentes au système à concevoir, en vue de proposer des solutions optimales et sûres. Le niveau de sûreté est quantifié par une probabilité de défaillance. Le problème d’optimisation consiste alors à s’assurer que cette probabilité reste inférieure à un seuil fixé par les donneurs d’ordres. La résolution de ce problème nécessite un grand nombre d’appels à la fonction d’état-limite caractérisant le problème de fiabilité sous-jacent. Ainsi,cette méthodologie devient complexe à appliquer dès lors que le dimensionnement s’appuie sur un modèle numérique coûteux à évaluer (e.g. un modèle aux éléments finis). Dans ce contexte, ce manuscrit propose une stratégie basée sur la substitution adaptative de la fonction d’état-limite par un méta-modèle par Krigeage. On s’est particulièrement employé à quantifier, réduire et finalement éliminer l’erreur commise par l’utilisation de ce méta-modèle en lieu et place du modèle original. La méthodologie proposée est appliquée au dimensionnement des coques géométriquement imparfaites soumises au flambement. / This thesis is a contribution to the resolution of the reliability-based design optimization problem. This probabilistic design approach is aimed at considering the uncertainty attached to the system of interest in order to provide optimal and safe solutions. The safety level is quantified in the form of a probability of failure. Then, the optimization problem consists in ensuring that this failure probability remains less than a threshold specified by the stakeholders. The resolution of this problem requires a high number of calls to the limit-state design function underlying the reliability analysis. Hence it becomes cumbersome when the limit-state function involves an expensive-to-evaluate numerical model (e.g. a finite element model). In this context, this manuscript proposes a surrogate-based strategy where the limit-state function is progressively replaced by a Kriging meta-model. A special interest has been given to quantifying, reducing and eventually eliminating the error introduced by the use of this meta-model instead of the original model. The proposed methodology is applied to the design of geometrically imperfect shells prone to buckling.

Nous proposons une approche spectrale permettant d'aborder des problèmes d'interpolation à noyaux dont la résolution numérique n'est pas directement envisageable. Un tel cas de figure se produit en particulier lorsque le nombre de données est infini. Nous considérons dans un premier temps le cadre de l'interpolation optimale dans les sous-espaces hilbertiens. Pour un problème donné, un opérateur intégral est défini à partir du noyau sous-jacent et d'une paramétrisation de l'ensemble des données basée sur un espace mesuré. La décomposition spectrale de l'opérateur est utilisée afin d'obtenir une formule de représentation pour l'interpolateur optimal et son approximation est alors rendu possible par troncature du spectre. Le choix de la mesure induit une fonction d'importance sur l'ensemble des données qui se traduit, en cas d'approximation, par une plus ou moins grande précision dans le rendu des données. Nous montrons à titre d'exemple comment cette approche peut être utilisée afin de rendre compte de contraintes de type "conditions aux limites" dans les modèles d'interpolation à noyaux. Le problème du conditionnement des processus gaussiens est également étudié dans ce contexte. Nous abordons enfin dans la dernière partie de notre manuscrit la notion de noyaux conditionnellement positifs. Nous proposons la définition générale de noyaux symétriques conditionnellement positifs relatifs à une espace de référence donné et développons la théorie des sous-espaces semi-hilbertiens leur étant associés. Nous étudions finalement la théorie de l'interpolation optimale dans cette classe d'espaces. / We propose a spectral approach for the resolution of kernel-based interpolation problems of which numerical solution can not be directly computed. Such a situation occurs in particular when the number of data is infinite. We first consider optimal interpolation in Hilbert subspaces. For a given problem, an integral operator is defined from the underlying kernel and a parameterization of the data set based on a measurable space. The spectral decomposition of the operator is used in order to obtain a representation formula for the optimal interpolator and spectral truncation allows its approximation. The choice of the measure on the parameters space introduces a hierarchy onto the data set which allows a tunable precision of the approximation. As an example, we show how this methodology can be used in order to enforce boundary conditions in kernel-based interpolation models. The Gaussian processes conditioning problem is also studied in this context. The last part of this thesis is devoted to the notion of conditionally positive kernels. We propose a general definition of symmetric conditionally positive kernels relative to a given space and exposed the associated theory of semi-Hilbert subspaces. We finally study the optimal interpolation problem in such spaces.

No description available.

The Hubble parameter H(z) is a measure of the expansion rate of the universe at redshift z. One method to determine it relies on inferring the slope of the redshift with respect to cosmic time, where galaxy ages can be used as a proxy for the latter. This method is used by Simon et al. in [1], where they present 8 determinations of the Hubble parameter. The results are surprisingly precise given the precision of their data set. Therefore, we reanalyze their data using three methods: chi-square minimization, Monte Carlo sampling, and Gaussian processes. The first two methods show that obtaining 8 independent values of the Hubble parameter yields significantly larger uncertainties than those presented by Simon et al. The last method yields a continuous inference of H(z) with lower uncertainties. However, this is obtained at the cost of having strong correlations, meaning that inferences at a wide range of redshifts provide essentially the same information. Furthermore, we demonstrate that obtaining 8 independent values for the Hubble parameter with the same precision as in [1] requires either significantly increasing the size of the data set, or significantly decreasing the uncertainty in the data. We conclude that their resulting Hubble parameter values can not be derived from the employed data. [1] J. Simon, L. Verde and R. Jimenez, Constraints on the redshift dependence of the dark energy potential, Physical Review D 71, 123001 (2005). / Hubbleparametern H(z) är ett mått på universums expansionshastighet vid rödskift z. En metod som bestämmer parametern bygger på att hitta lutningen av sambandet mellan rödskift och kosmisk tid, där det sistnämnda går att ersätta med galaxåldrar. Denna metod används av Simon et al. i [1], där de presenterar 8 värden av Hubbleparametern. Resultaten är förvånansvärt precisa, med tanke på precisionen i deras data. Vi omanalyserar därför deras data med tre metoder: chi-2-miniminering, Monte Carlo-sampling och Gaussiska processer. De två första metoderna visar att när 8 oberoende värden av Hubbleparametern bestäms fås mycket större osäkerheter än de som presenteras av Simon et al. Den sistnämnda metoden ger en kontinuerlig funktion H(z) med lägre osäkerheter. Priset för detta är dock starka korrelationer, det vill säga att resultat vid många olika rödskift innehåller i princip samma information. Utöver detta visar vi att det krävs antingen en mycket större mängd data eller mycket mindre osäkerheter i datan för att kunna bestämma 8 oberoende värden av Hubbleparametern med samma precision som i [1]. Vi drar slutsatsen att deras värden av Hubbleparametern inte kan fås med den data som använts. [1] J. Simon, L. Verde and R. Jimenez, Constraints on the redshift dependence of the dark energy potential, Physical Review D 71, 123001 (2005).

En los ultimos años el número de casos de ceguera se ha reducido significativamente. A pesar de este hecho, la Organización Mundial de la Salud estima que un 80% de los casos de pérdida de visión (285 millones en 2010) pueden ser evitados si se diagnostican en sus estadios más tempranos y son tratados de forma efectiva. Para cumplir esta propuesta se pretende que los servicios de atención primaria incluyan un seguimiento oftalmológico de sus pacientes así como fomentar campañas de cribado en centros proclives a reunir personas de alto riesgo. Sin embargo, estas soluciones exigen una alta carga de trabajo de personal experto entrenado en el análisis de los patrones anómalos propios de cada enfermedad. Por lo tanto, el desarrollo de algoritmos para la creación de sistemas de cribado automáticos juga un papel vital en este campo. La presente tesis persigue la identificacion automática del daño retiniano provocado por dos de las patologías más comunes en la sociedad actual: la retinopatía diabética (RD) y la degenaración macular asociada a la edad (DMAE). Concretamente, el objetivo final de este trabajo es el desarrollo de métodos novedosos basados en la extracción de características de la imagen de fondo de ojo y clasificación para discernir entre tejido sano y patológico. Además, en este documento se proponen algoritmos de pre-procesado con el objetivo de normalizar la alta variabilidad existente en las bases de datos publicas de imagen de fondo de ojo y eliminar la contribución de ciertas estructuras retinianas que afectan negativamente en la detección del daño retiniano. A diferencia de la mayoría de los trabajos existentes en el estado del arte sobre detección de patologías en imagen de fondo de ojo, los métodos propuestos a lo largo de este manuscrito evitan la necesidad de segmentación de las lesiones o la generación de un mapa de candidatos antes de la fase de clasificación. En este trabajo, Local binary patterns, perfiles granulométricos y la dimensión fractal se aplican de manera local para extraer información de textura, morfología y tortuosidad de la imagen de fondo de ojo. Posteriormente, esta información se combina de diversos modos formando vectores de características con los que se entrenan avanzados métodos de clasificación formulados para discriminar de manera óptima entre exudados, microaneurismas, hemorragias y tejido sano. Mediante diversos experimentos, se valida la habilidad del sistema propuesto para identificar los signos más comunes de la RD y DMAE. Para ello se emplean bases de datos públicas con un alto grado de variabilidad sin exlcuir ninguna imagen. Además, la presente tesis también cubre aspectos básicos del paradigma de deep learning. Concretamente, se presenta un novedoso método basado en redes neuronales convolucionales (CNNs). La técnica de transferencia de conocimiento se aplica mediante el fine-tuning de las arquitecturas de CNNs más importantes en el estado del arte. La detección y localización de exudados mediante redes neuronales se lleva a cabo en los dos últimos experimentos de esta tesis doctoral. Cabe destacar que los resultados obtenidos mediante la extracción de características "manual" y posterior clasificación se comparan de forma objetiva con las predicciones obtenidas por el mejor modelo basado en CNNs. Los prometedores resultados obtenidos en esta tesis y el bajo coste y portabilidad de las cámaras de adquisión de imagen de retina podrían facilitar la incorporación de los algoritmos desarrollados en este trabajo en un sistema de cribado automático que ayude a los especialistas en la detección de patrones anomálos característicos de las dos enfermedades bajo estudio: RD y DMAE. / In last years, the number of blindness cases has been significantly reduced. Despite this promising news, the World Health Organisation estimates that 80% of visual impairment (285 million cases in 2010) could be avoided if diagnosed and treated early. To accomplish this purpose, eye care services need to be established in primary health and screening campaigns should be a common task in centres with people at risk. However, these solutions entail a high workload for trained experts in the analysis of the anomalous patterns of each eye disease. Therefore, the development of algorithms for automatic screening system plays a vital role in this field. This thesis focuses on the automatic identification of the retinal damage provoked by two of the most common pathologies in the current society: diabetic retinopathy (DR) and age-related macular degeneration (AMD). Specifically, the final goal of this work is to develop novel methods, based on fundus image description and classification, to characterise the healthy and abnormal tissue in the retina background. In addition, pre-processing algorithms are proposed with the aim of normalising the high variability of fundus images and removing the contribution of some retinal structures that could hinder in the retinal damage detection. In contrast to the most of the state-of-the-art works in damage detection using fundus images, the methods proposed throughout this manuscript avoid the necessity of lesion segmentation or the candidate map generation before the classification stage. Local binary patterns, granulometric profiles and fractal dimension are locally computed to extract texture, morphological and roughness information from retinal images. Different combinations of this information feed advanced classification algorithms formulated to optimally discriminate exudates, microaneurysms, haemorrhages and healthy tissues. Through several experiments, the ability of the proposed system to identify DR and AMD signs is validated using different public databases with a large degree of variability and without image exclusion. Moreover, this thesis covers the basics of the deep learning paradigm. In particular, a novel approach based on convolutional neural networks is explored. The transfer learning technique is applied to fine-tune the most important state-of-the-art CNN architectures. Exudate detection and localisation tasks using neural networks are carried out in the last two experiments of this thesis. An objective comparison between the hand-crafted feature extraction and classification process and the prediction models based on CNNs is established. The promising results of this PhD thesis and the affordable cost and portability of retinal cameras could facilitate the further incorporation of the developed algorithms in a computer-aided diagnosis (CAD) system to help specialists in the accurate detection of anomalous patterns characteristic of the two diseases under study: DR and AMD. / En els últims anys el nombre de casos de ceguera s'ha reduït significativament. A pesar d'este fet, l'Organització Mundial de la Salut estima que un 80% dels casos de pèrdua de visió (285 milions en 2010) poden ser evitats si es diagnostiquen en els seus estadis més primerencs i són tractats de forma efectiva. Per a complir esta proposta es pretén que els servicis d'atenció primària incloguen un seguiment oftalmològic dels seus pacients així com fomentar campanyes de garbellament en centres regentats per persones d'alt risc. No obstant això, estes solucions exigixen una alta càrrega de treball de personal expert entrenat en l'anàlisi dels patrons anòmals propis de cada malaltia. Per tant, el desenrotllament d'algoritmes per a la creació de sistemes de garbellament automàtics juga un paper vital en este camp. La present tesi perseguix la identificació automàtica del dany retiniano provocat per dos de les patologies més comunes en la societat actual: la retinopatia diabètica (RD) i la degenaración macular associada a l'edat (DMAE) . Concretament, l'objectiu final d'este treball és el desenrotllament de mètodes novedodos basats en l'extracció de característiques de la imatge de fons d'ull i classificació per a discernir entre teixit sa i patològic. A més, en este document es proposen algoritmes de pre- processat amb l'objectiu de normalitzar l'alta variabilitat existent en les bases de dades publiques d'imatge de fons d'ull i eliminar la contribució de certes estructures retinianas que afecten negativament en la detecció del dany retiniano. A diferència de la majoria dels treballs existents en l'estat de l'art sobre detecció de patologies en imatge de fons d'ull, els mètodes proposats al llarg d'este manuscrit eviten la necessitat de segmentació de les lesions o la generació d'un mapa de candidats abans de la fase de classificació. En este treball, Local binary patterns, perfils granulometrics i la dimensió fractal s'apliquen de manera local per a extraure informació de textura, morfologia i tortuositat de la imatge de fons d'ull. Posteriorment, esta informació es combina de diversos modes formant vectors de característiques amb els que s'entrenen avançats mètodes de classificació formulats per a discriminar de manera òptima entre exsudats, microaneurismes, hemorràgies i teixit sa. Per mitjà de diversos experiments, es valida l'habilitat del sistema proposat per a identificar els signes més comuns de la RD i DMAE. Per a això s'empren bases de dades públiques amb un alt grau de variabilitat sense exlcuir cap imatge. A més, la present tesi també cobrix aspectes bàsics del paradigma de deep learning. Concretament, es presenta un nou mètode basat en xarxes neuronals convolucionales (CNNs) . La tècnica de transferencia de coneixement s'aplica per mitjà del fine-tuning de les arquitectures de CNNs més importants en l'estat de l'art. La detecció i localització d'exudats per mitjà de xarxes neuronals es du a terme en els dos últims experiments d'esta tesi doctoral. Cal destacar que els resultats obtinguts per mitjà de l'extracció de característiques "manual" i posterior classificació es comparen de forma objectiva amb les prediccions obtingudes pel millor model basat en CNNs. Els prometedors resultats obtinguts en esta tesi i el baix cost i portabilitat de les cambres d'adquisión d'imatge de retina podrien facilitar la incorporació dels algoritmes desenrotllats en este treball en un sistema de garbellament automàtic que ajude als especialistes en la detecció de patrons anomálos característics de les dos malalties baix estudi: RD i DMAE. / Colomer Granero, A. (2018). Fundus image analysis for automatic screening of ophthalmic pathologies [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/99745