No / An analytic framework is devised, based on the principle of maximum entropy (ME), for the performance modelling and evaluation of a wireless GSM/GPRS cell supporting bursty multiple class traffic of voice calls and data packets under complete partitioning (CPS), partial sharing (PSS) and aggregate sharing (ASS) traffic handling schemes. Three distinct open queueing network models (QNMS) under CPS, PSS and ASS, respectively, are described, subject to external compound Poisson traffic processes and generalised exponential (GE) transmission times under a repetitive service blocking mechanism and a complete buffer sharing management rule. Each QNM generally consists of three building block stations, namely a loss system with GSM/GPRS traffic and a system of access and transfer finite capacity queues in tandem dealing with GPRS traffic under head-of-line and discriminatory processor sharing scheduling disciplines, respectively. The analytic methodology is illustrated by focusing on the performance study of the GE-type tandem queueing system for GPRS under a CPS. An ME product-form approximation is characterised leading into a decomposition of the tandem system into individual queues and closed-form ME expressions for state and blocking probabilities are presented. Typical numerical examples are included to validate the ME solutions against simulation and study the effect of external GPRS bursty traffic upon the performance of the cell. Moreover, an overview of recent extensions of the work towards the analysis of a GE-type multiple server finite capacity queue with preemptive resume priorities and its implications towards the performance modelling and evaluation of GSM/GPRS cells with PSS and ASS are included. / ,

Remote Sensing (RS) technology provides valuable information about Earth by enabling an overview of the planet from above, making it a much-needed resource for many applications. Given the abundance of RS data and continued urbanisation, there is a need for efficient approaches to leverage RS data and its unique characteristics for the assessment and management of urban areas. Consequently, employing Deep Learning (DL) for RS applications has attracted much attention over the past few years. In this thesis, novel datasets consisting of satellite RS images over urban areas in Sweden were compiled from Sentinel-2 multispectral, Sentinel-1 Synthetic Aperture Radar (SAR) and Urban Atlas 2018 Land Use/Land Cover (LULC) data. Then, DL was applied for multiband and multiclass semantic image segmentation of LULC. The contributions of complementary spectral, temporal and SAR data and spectral indices to LULC classification performance compared to using only Sentinel-2 data with red, green and blue spectral bands were investigated by implementing DL models based on the fully convolutional network-based architecture, U-Net, and performing data fusion. Promising results were achieved with 25 possible LULC classes. Furthermore, almost all DL models at an overall model level and all DL models at an individual class level for most LULC classes benefited from complementary satellite RS data with varying degrees of classification improvement. Additionally, practical knowledge and insights were gained from evaluating the results and are presented regarding satellite RS data characteristics and semantic segmentation of LULC in urban areas. The obtained results are helpful for practitioners and researchers applying or intending to apply DL for semantic segmentation of LULC in general and specifically in Swedish urban environments.

Les avancées technologiques ont permis le stockage de grandes masses de données en termes de taille (nombre d’observations) et de dimensions (nombre de variables).Ces données nécessitent de nouvelles méthodes, notamment en modélisation prédictive (data science ou science des données), de traitement statistique adaptées à leurs caractéristiques. Dans le cadre de cette thèse, nous nous intéressons plus particulièrement aux données dont le nombre de variables est élevé comparé au nombre d’observations.Pour ces données, une réduction du nombre de variables initiales, donc de dimensions, par la sélection d’un sous-ensemble optimal, s’avère nécessaire, voire indispensable.Elle permet de réduire la complexité, de comprendre la structure des données et d’améliorer l’interprétation des résultats et les performances du modèle de prédiction ou de classement en éliminant les variables bruit et/ou redondantes.Nous nous intéressons plus précisément à la sélection de variables dans le cadre de l’apprentissage supervisé et plus spécifiquement de la discrimination à catégories multiples dite multi-classes. L’objectif est de proposer de nouvelles méthodes de sélection de variables pour les modèles de discrimination multi-classes appelés Machines à Vecteurs Support Multiclasses (MSVM).Deux approches sont proposées dans ce travail. La première, présentée dans un contexte classique, consiste à sélectionner le sous-ensemble optimal de variables en utilisant le critère de "la borne rayon marge" majorante du risque de généralisation des MSVM. Quant à la deuxième approche, elle s’inscrit dans un contexte topologique et utilise la notion de graphes de voisinage et le critère de degré d’équivalence topologique en discrimination pour identifier les variables pertinentes qui constituent le sous-ensemble optimal du modèle MSVM.L’évaluation de ces deux approches sur des données simulées et d’autres réelles montre qu’elles permettent de sélectionner, à partir d’un grand nombre de variables initiales, un nombre réduit de variables explicatives avec des performances similaires ou encore meilleures que celles obtenues par des méthodes concurrentes. / The technological progress has allowed the storage of large amounts of data in terms of size (number of observations) and dimensions (number of variables). These data require new methods, especially for predictive modeling (data science), of statistical processing adapted to their characteristics. In this thesis, we are particularly interested in the data with large numberof variables compared to the number of observations.For these data, reducing the number of initial variables, hence dimensions, by selecting an optimal subset is necessary, even imperative. It reduces the complexity, helps to understand the data structure, improves the interpretation of the results and especially enhances the performance of the forecasting model by eliminating redundant and / or noise variables.More precisely, we are interested in the selection of variables in the context of supervised learning, specifically of multiclass discrimination. The objective is to propose some new methods of variable selection for multiclass discriminant models called Multiclass Support Vector Machines (MSVM).Two approaches are proposed in this work. The first one, presented in a classical context, consist in selecting the optimal subset of variables using the radius margin upper bound of the generalization error of MSVM. The second one, proposed in a topological context, uses the concepts of neighborhood graphs and the degree of topological equivalence in discriminationto identify the relevant variables and to select the optimal subset for an MSVM model.The evaluation of these two approaches on simulated and real data shows that they can select from a large number of initial variables, a reduced number providing equal or better performance than those obtained by competing methods.

In this thesis, we investigate the problem of human layout estimation in unconstrained still images. This involves predicting the spatial configuration of body parts. We start our investigation with pictorial structure models and propose an efficient method of model fitting using skin regions. To detect the skin, we learn a colour model locally from the image by detecting the facial region. The resulting skin detections are also used for hand localisation. Our next contribution is a comprehensive dataset of 2D hand images. We collected this dataset from publicly available image sources, and annotated images with hand bounding boxes. The bounding boxes are not axis aligned, but are rather oriented with respect to the wrist. Our dataset is quite exhaustive as it includes images of different hand shapes and layout configurations. Using our dataset, we train a hand detector that is robust to background clutter and lighting variations. Our hand detector is implemented as a two-stage system. The first stage involves proposing hand hypotheses using complementary image features, which are then evaluated by the second stage classifier. This improves both precision and recall and results in a state-of-the-art hand detection method. In addition we develop a new method of non-maximum suppression based on super-pixels. We also contribute an efficient training algorithm for structured output ranking. In our algorithm, we reduce the time complexity of an expensive training component from quadratic to linear. This algorithm has a broad applicability and we use it for solving human layout estimation and taxonomic multiclass classification problems. For human layout, we use different body part detectors to propose part candidates. These candidates are then combined and scored using our ranking algorithm. By applying this bottom-up approach, we achieve accurate human layout estimation despite variations in viewpoint and layout configuration. In the multiclass classification problem, we define the misclassification error using a class taxonomy. The problem then reduces to a structured output ranking problem and we use our ranking method to optimise it. This allows inclusion of semantic knowledge about the classes and results in a more meaningful classification system. Lastly, we substantiate our ranking algorithm with theoretical proofs and derive the generalisation bounds for it. These bounds prove that the training error reduces to the lowest possible error asymptotically.

Multilabel classification has attracted much interest in recent times due to the wide applicability of the problem and the challenges involved in learning a classifier for multilabeled data. A crucial aspect of multilabel classification is to discover the structure and order of correlations among labels and their effect on the quality of the classifier. In this work, we propose a structural Support Vector Machine (structural SVM) based framework which enables us to systematically investigate the importance of label correlations in multi-label classification. The proposed framework is very flexible and provides a unified approach to handle multiple correlation orders and structures in an adaptive manner and helps to effectively assess the importance of label correlations in improving the generalization performance. We perform extensive empirical evaluation on several datasets from different domains and present results on various performance metrics. Our experiments provide for the first time, interesting insights into the following questions: a) Are label correlations always beneficial in multilabel classification? b) What effect do label correlations have on multiple performance metrics typically used in multilabel classification? c) Is label correlation order significant and if so, what would be the favorable correlation order for a given dataset and a given performance metric? and d) Can we make useful suggestions on the label correlation structure?

Made available in DSpace on 2016-06-02T19:05:36Z (GMT). No. of bitstreams: 1 2598.pdf: 3349204 bytes, checksum: 01649491fd1f03aa5a11b9191727f88b (MD5) Previous issue date: 2009-06-29 / Financiadora de Estudos e Projetos / This master thesis describes an intelligent classifier system applied to multiclass non-linearly separable problems called Slicer. The system adopts a low computacional cost supervised learning strategy (evaluated as ) based on linear separability. During the learning period the system determines a set of hyperplanes associated to oneclass regions (sub-spaces). In classification tasks the classifier system uses the hyperplanes as a set of if-then-else rules to infer the class of the input attribute vector (non classified object). Among other characteristics, the intelligent classifier system is able to: deal with missing attribute values examples; reject noise examples during learning; adjust hyperplane parameters to improve the definition of the one-class regions; and eliminate redundant rules. The fuzzy theory is considered to design a hybrid version with features such as approximate reasoning and parallel inference computation. Different classification methods and benchmarks are considered for evaluation. The classifier system Slicer reaches acceptable results in terms of accuracy, justifying future investigation effort. / Este trabalho de mestrado descreve um sistema classificador inteligente aplicado a problemas multiclasse não-linearmente separáveis chamado Slicer. O sistema adota uma estratégia de aprendizado supervisionado de baixo custo computacional (avaliado em ) baseado em separabilidade linear. Durante o período de aprendizagem o sistema determina um conjunto de hiperplanos associados a regiões de classe única (subespaços). Nas tarefas de classificação o sistema classificador usa os hiperplanos como um conjunto de regras se-entao-senao para inferir a classe do vetor de atributos dado como entrada (objeto a ser classificado). Entre outras caracteristicas, o sistema classificador é capaz de: tratar atributos faltantes; eliminar ruídos durante o aprendizado; ajustar os parâmetros dos hiperplanos para obter melhores regiões de classe única; e eliminar regras redundantes. A teoria nebulosa é considerada para desenvolver uma versão híbrida com características como raciocínio aproximado e simultaneidade no mecanismo de inferência. Diferentes métodos de classificação e domínios são considerados para avaliação. O sistema classificador Slicer alcança resultados aceitáveis em termos de acurácia, justificando investir em futuras investigações.

The problem of feature selection is critical in several areas of machine learning and data analysis such as, for example, cancer classification using gene expression data, text categorization, etc. In this work, we consider feature selection for supervised learning problems, where one wishes to select a small set of features that facilitate learning a good prediction model in the reduced feature space. Our interest is primarily in filter methods that select features independently of the learning algorithm to be used and are generally faster to implement compared to other types of feature selection algorithms. Many common filter methods for feature selection make use of information-theoretic criteria such as those based on mutual information to guide their search process. However, even in simple binary classification problems, mutual information based methods do not always select the best set of features in terms of the Bayes error. In this thesis, we develop a general approach for selecting a set of features that directly aims to minimize the Bayes error in the reduced feature space with respect to the loss or performance measure of interest. We show that the mutual information based criterion is a special case of our setting when the loss function of interest is the logarithmic loss for class probability estimation. We give a greedy forward algorithm for approximately optimizing this criterion and demonstrate its application to several supervised learning problems including binary classification (with 0-1 error, cost-sensitive error, and F-measure), binary class probability estimation (with logarithmic loss), bipartite ranking (with pairwise disagreement loss), and multiclass classification (with multiclass 0-1 error). Our experiments suggest that the proposed approach is competitive with several state-of-the art methods.

Les travaux présentés dans cette thèse relèvent de l'étude des méthodes de classification linéaires, c'est à dire l'étude de méthodes ayant pour but la catégorisation de données en différents groupes à partir d'un jeu d'exemples, préalablement étiquetés, disponible en amont et appelés ensemble d'apprentissage. En pratique, l'acquisition d'un tel ensemble d'apprentissage peut être difficile et/ou couteux, la catégorisation d'un exemple étant de fait plus ardu que l'obtention de dudit exemple. Cette disparité entre la disponibilité des données et notre capacité à constituer un ensemble d'apprentissage étiqueté a été un des problèmes centraux de l'apprentissage automatique et ce manuscrit s’intéresse à deux solutions usuellement considérées pour contourner ce problème : l'apprentissage en présence de données bruitées et l'apprentissage actif. / The works presented in this thesis fall within the general framework of linear classification, that is the problem of categorizing data into two or more classes based on on a training set of labelled data. In practice though acquiring labeled examples might prove challenging and/or costly as data are inherently easier to obtain than to label. Dealing with label scarceness have been a motivational goal in the machine learning literature and this work discuss two settings related to this problem: learning in the presence of noise and active learning.

Scheduling policies adopted for statistical multiplexing should provide delay differentiation between different traffic classes, where each class represents an aggregate traffic of individual applications having same target-queueing-delay requirements. We propose scheduling to optimally balance weighted mean instanteneous queue lengths and later weighted mean cumulative queue lengths as an approach to delay differentiation, where the class weights are set inversely proportional to the respective products of target delays and packet arrival rates. In particular, we assume a discrete-time, two-class, single-server queueing model with unit service time per packet and provide mathematical frame-work throughout our work. For iid Bernoulli packet arrivals, using a step-wise cost-dominance analytical approach using instantaneous queue lengths alone, for a class of one-stage cost functions not necessarily convex, we find the structure of the total-cost optimal policies for a part of the state space. We then consider two particular one-stage cost functions for finding two scheduling policies that are total-cost optimal for the whole state-space. The policy for the absolute weighted difference cost function minimizes the stationary mean, and the policy for the weighted sum-of-square cost function minimizes the stationary second-order moment, of the absolute value of the weighted difference of queue lengths. For the case of weighted sum-of-square cost function, the ‘iid Bernoulli arrivals’ assumption can be relaxed to either ‘iid arrivals with general batch sizes’ or to ‘Markovian zero-one arrivals’ for all of the state space, but for the linear switching curve. We then show that the average cost, starting from any initial state, exists, and is finite for every stationary work-conserving policy for our choices of the one-stage cost-function. This is shown for arbitrary number of class queues and for any i.i.d. batch arrival processes with finite appropriate moments. We then use cumulative queue lengths information in the one-step cost function of the optimization formulation and obtain an optimal myopic policy with 3 stages to go for iid arrivals with general batch sizes. We show analytically that this policy achieves the given target delay ratio in the long run under finite buffer assumption, given that feasibility conditions are satisfied. We take recourse to numerical value iteration to show the existence of average-cost for this policy. Simulations with varied class-weights for Bernoulli arrivals and batch arrivals with Poisson batch sizes show that this policy achieves mean queueing delays closer to the respective target delays than the policy obtained earlier. We also note that the coefficients of variation of the queueing delays of both the classes using cumulative queue lengths are of the same order as those using instantaneous queue lengths. Moreover, the short-term behaviour of the optimal myopic policy using cumulative queue lengths is superior to the existing standard policy reported by Coffman and Mitrani by a factor in the range of 3 to 8. Though our policy performs marginally poorer compared to the value-iterated, sampled, and then stationarily employed policy, the later lacks any closed-form structure. We then modify the definition of the third state variable and look to directly balance weighted mean delays. We come up with another optimal myopic policy with 3 stages to go, following which the error in the ratio of mean delays decreases as the window-size, as opposed to the policy mentioned in the last paragraph, wherein the error decreases as the square-root of the window-size. We perform numerical value-iteration to show the existence of average-cost and study the performance by simulation. Performance of our policy is comparable with the value-iterated, sampled, and then stationarily employed policy, reported by Mallesh. We have then studied general inter-arrival time processes and obtained the optimal myopic policy for the Pareto inter-arrival process, in particular. We have supported with simulation that our policy fares similarly to the PAD policy, reported by Dovrolis et. al., which is primarily heuristic in nature. We then model the possible packet errors in the multiplexed channel by either a Bernoulli process, or a Markov modulated Bernoulli process with two possible channel states. We also consider two possible round-trip-time values for control information, namely zero and one-slot. The policies that are next-stage optimal (for zero round-trip-time), and two-stage optimal (for one-slot round-trip-time) are obtained. Simulations with varied class-weights for Bernoulli arrivals and batch arrivals with Poisson batch sizes show that these policies indeed achieve mean queueing delays very close to the respective target delays. We also obtain the structure for optimal policies with N = 2 + ⌈rtt⌉ stages-to-go for generic values of rtt, and which need not be multiple of time-slots.

La première partie de cette thèse introduit de nouveaux algorithmes de décomposition parcimonieuse de signaux. Basés sur Matching Pursuit (MP) ils répondent au problème suivant : comment réduire le temps de calcul de l'étape de sélection de MP, souvent très coûteuse. En réponse, nous sous-échantillonnons le dictionnaire à chaque itération, en lignes et en colonnes. Nous montrons que cette approche fondée théoriquement affiche de bons résultats en pratique. Nous proposons ensuite un algorithme itératif de descente de gradient par blocs de coordonnées pour sélectionner des caractéristiques en classification multi-classes. Celui-ci s'appuie sur l'utilisation de codes correcteurs d'erreurs transformant le problème en un problème de représentation parcimonieuse simultanée de signaux. La deuxième partie expose de nouvelles inégalités de concentration empiriques de type Bernstein. En premier, elles concernent la théorie des U-statistiques et sont utilisées pour élaborer des bornes en généralisation dans le cadre d'algorithmes de ranking. Ces bornes tirent parti d'un estimateur de variance pour lequel nous proposons un algorithme de calcul efficace. Ensuite, nous présentons une version empirique de l'inégalité de type Bernstein proposée par Freedman [1975] pour les martingales. Ici encore, la force de notre borne réside dans l'introduction d'un estimateur de variance calculable à partir des données. Cela nous permet de proposer des bornes en généralisation pour l'ensemble des algorithmes d'apprentissage en ligne améliorant l'état de l'art et ouvrant la porte à une nouvelle famille d'algorithmes d'apprentissage tirant parti de cette information empirique. / The first part of this thesis introduces new algorithms for the sparse encoding of signals. Based on Matching Pursuit (MP) they focus on the following problem : how to reduce the computation time of the selection step of MP. As an answer, we sub-sample the dictionary in line and column at each iteration. We show that this theoretically grounded approach has good empirical performances. We then propose a bloc coordinate gradient descent algorithm for feature selection problems in the multiclass classification setting. Thanks to the use of error-correcting output codes, this task can be seen as a simultaneous sparse encoding of signals problem. The second part exposes new empirical Bernstein inequalities. Firstly, they concern the theory of the U-Statistics and are applied in order to design generalization bounds for ranking algorithms. These bounds take advantage of a variance estimator and we propose an efficient algorithm to compute it. Then, we present an empirical version of the Bernstein type inequality for martingales by Freedman [1975]. Again, the strength of our result lies in the variance estimator computable from the data. This allows us to propose generalization bounds for online learning algorithms which improve the state of the art and pave the way to a new family of learning algorithms taking advantage of this empirical information.