Landminor är ett stort problem både under och efter krigstid. De metoder som används för att detektera minor har inte ändrats mycket sedan 1940-talet. Forskning med mål att utvärdera olika elektro-optiska sensorer och metoder som skulle kunna användas för att skapa mer effektiv min-detektion genomförs på FOI. Försök som har gjorts med data från bland annat laser-radar och IR-sensorer har gett intressanta resultat. I det här examensarbetet utvärderades olika fenomen och egenskaper i laser-radar- och IR-data. De testade egenskaperna var intensitet, IR, ytlikhet och höjd. En metod som segmenterar intressanta objekt och bakgrundsdata utformades och implementerades. Metoden använde sig av expectation-maximization-skattning och ett minimum message length-kriterium. Ett scatter separability-kriterium användes för att bestämma kvalitén på de olika egenskaperna och på den resulterande segmenteringen. Data insamlad under en mätkampanj av FOI användes för att testa metoden. Resultatet visade bland annat att ytlikhetsmåttet gav en bra segmentering för stora objekt med släta ytor, men var sämre för små objekt med skrovliga ytor. Vid jämförelse med en manuellt skapad mål-mask visade det sig att metoden klarade av att välja ut egenskaper som i många fall gav en godkänd segmentering. / Land mines are a huge problem in conflict time and after. Methods used to detect mines have not changed much since the 1940's. Research aiming to evaluate output from different electro-optical sensors and develop methods for more efficient mine detection is performed at FOI. Early experiments with laser radar sensors show promising results, as do analysis of data from infrared sensors. In this thesis, an evaluation is made of features found in laser radar- and in infrared -sensor data. The tested features are intensity, infrared, a surfaceness feature extracted from the laser radar data and height above an estimated ground plane. A method for segmenting interesting objects from background data using theexpectation-maximization algorithm and a minimum message length criterion is designed and implemented. A scatter separability criterion is utilized to determine the quality of the features and the resulting segmentation. The method is tested on real data from a field trial performed by FOI. The results show that the surfaceness feature supports the segmentation of larger object with smooth surfaces but gives no contribution to small object with irregular surfaces. The method produces a decent result of selecting contributing features for different neighbourhoods of a scene. A comparison with a manually created target mask of the neighbourhood and the segmented components show that in most cases a high percentage separation of mine data and background data is possible.

Virtually all methods of learning dynamic models from data start from the same basic assumption: that the learning algorithm will be provided with a single or multiple sequences of data generated from the dynamic model. However, in quite a few modern time series modeling tasks, the collection of reliable time series data turns out to be a major challenge, due to either slow progression of the dynamic process of interest, or inaccessibility of repetitive measurements of the same dynamic process over time. In most of those situations, however, we observe that it is easier to collect a large amount of non-sequence samples, or random snapshots of the dynamic process of interest without time information. This thesis aims to exploit such non-sequence data in learning a few widely used dynamic models, including fully observable, linear and nonlinear models as well as Hidden Markov Models (HMMs). For fully observable models, we point out several issues on model identifiability when learning from non-sequence data, and develop EM-type learning algorithms based on maximizing approximate likelihood. We also consider the setting where a small amount of sequence data are available in addition to non-sequence data, and propose a novel penalized least square approach that uses non-sequence data to regularize the model. For HMMs, we draw inspiration from recent advances in spectral learning of latent variable models and propose spectral algorithms that provably recover the model parameters, under reasonable assumptions on the generative process of non-sequence data and the true model. To the best of our knowledge, this is the first formal guarantee on learning dynamic models from non-sequence data. We also consider the case where little sequence data are available, and propose learning algorithms that, as in the fully observable case, use non-sequence data to provide regularization, but does so in combination with spectral methods. Experiments on synthetic data and several real data sets, including gene expression and cell image time series, demonstrate the effectiveness of our proposed methods. In the last part of the thesis we return to the usual setting of learning from sequence data, and consider learning bi-clustered vector auto-regressive models, whose transition matrix is both sparse, revealing significant interactions among variables, and bi-clustered, identifying groups of variables that have similar interactions with other variables. Such structures may aid other learning tasks in the same domain that have abundant non-sequence data by providing better regularization in our proposed non-sequence methods.

With the growing popularity of business analytics, companies experience an increasing need of reliable data. Although the availability of behavioural data showing what the consumers do has increased, the access to data showing consumer mentality, what the con- sumers actually think, remain heavily dependent on tracking surveys. This thesis inves- tigates the performance of a Dynamic Factor Model using respondent-level data gathered through repeated cross-sectional surveys. Through Monte Carlo simulations, the model was shown to improve the accuracy of brand tracking estimates by double digit percent- ages, or equivalently reducing the required amount of data by more than a factor 2, while maintaining the same level of accuracy. Furthermore, the study showed clear indications that even greater performance benefits are possible.

Adverse drug reactions are unwanted effects alongside the intended benefit of a drug and might be responsible for 3-7\% of hospitalizations. Finding such reactions is partly done by analysing individual case safety reports (ICSR) of adverse events. The reports consist of categorical terms that describe the event.Data-driven identification of suspected adverse drug reactions using this data typically considers single adverse event terms, one at a time. This single term approach narrows the identification of reports and information in the reports is ignored during the search. If one instead assumes that each report is connected to a topic, then by creating a cluster of the reports that are connected to the topic more reports would be identified. More context would also be provided by virtue of the topics. This thesis takes place at Uppsala Monitoring Centre which has implemented a probabilistic model of how an ICSR, and its topic, is assumed to be generated. The parameters of the model are estimated with expectation maximization (EM), which also assigns the reports to clusters. The clusters are improved with Consensus Clustering that identify groups of reports that tend to be grouped together by several runs of EM. Additionally, in order to not cluster outlying reports all clusters below a certain size are excluded. The objective of the thesis is to improve the algorithm in terms of computational efficiency and quality, as measured by stability and clinical coherence. The convergence of EM is improved using stepwise EM, which resulted in a speed up of at least 1.4, and a decrease of the computational complexity. With all the speed improvements the speed up factor of the entire algorithm can reach 2 but is constrained by the size of the data. In order to improve the clusters' quality, the community detection algorithm Leiden is used. It is able to improve the stability with the added benefit of increasing the number of clustered reports. The clinical coherence score performs worse with Leiden. There are good reasons to further investigate the benefits of Leiden as there were suggestions that community detection identified clusters with greater resolution that still appeared clinically coherent in a posthoc analysis.

This thesis presents a statistical method to identify the test escapes. Test often acquires parametric measurements as a function of logical state of a chip. The usual method of classifying chips as pass or fail is to compare each state measurement to a test limit. Subtle manufacturing defects are escaping the test limits due to process variations in deep sub-micron technologies which results in mixing of healthy and faulty parametric test measurements. This thesis identifies the chips with subtle defects by using rank order of the parametric measurements. A hypothesis is developed that a defect is likely to disturb the defect-free ranking, whereas a shift caused by process variations will not affect the rank. The hypothesis does not depend on a-priori knowledge of a defect-free ranking of parametric measurements. This thesis introduces a modified Estimation Maximization (EM) algorithm to separate the healthy and faulty tau components calculated from parametric responses of die pairs on a wafer. The modified EM uses generalized beta distributions to model the two components of tau mixture distribution. The modified EM estimates the faulty probability of each die on a wafer. The sensitivity of the modified EM is evaluated using Monte Carlo simulations. The modified EM is applied on production product A. An average 30% reduction in DPPM (defective parts per million) is observed in Product A across all lots.

No description available.

RNA-Seq is an increasingly popular approach to transcriptome profiling that uses the capabilities of next generation sequencing technologies and provides better measurement of levels of transcripts and their isoforms. In this thesis, we apply RNA-Seq protocol and transcriptome quantification to estimate gene expression and pathway activity levels. We present a novel method, called IsoDE, for differential gene expression analysis based on bootstrapping. In the first version of IsoDE, we compared the tool against four existing methods: Fisher's exact test, GFOLD, edgeR and Cuffdiff on RNA-Seq datasets generated using three different sequencing technologies, both with and without replicates. We also introduce the second version of IsoDE which runs 10 times faster than the first implementation due to some in-memory processing applied to the underlying gene expression frequencies estimation tool and we also perform more optimization on the analysis. The second part of this thesis presents a set of tools to differentially analyze metabolic pathways from RNA-Seq data. Metabolic pathways are series of chemical reactions occurring within a cell. We focus on two main problems in metabolic pathways differential analysis, namely, differential analysis of their inferred activity level and of their estimated abundance. We validate our approaches through differential expression analysis at the transcripts and genes levels and also through real-time quantitative PCR experiments. In part Four, we present the different packages created or updated in the course of this study. We conclude with our future work plans for further improving IsoDE 2.0.

La séparation de sources avec un seul capteur est un problème très récent, qui attire de plus en plus d'attention dans le monde scientifique. Cependant, il est loin d'être résolu et, même plus, il ne peut pas être résolu en toute généralité. La difficulté principale est que, ce problème étant extrêmement sous déterminé, il faut disposer de fortes connaissances sur les sources pour pouvoir les séparer. Pour une grande partie des méthodes de séparation, ces connaissances sont représentées par des modèles statistiques des sources, notamment par des Modèles de Mélange de Gaussiennes (MMG), qui sont appris auparavant à partir d'exemples. L'objet de cette thèse est d'étudier les méthodes de séparation basées sur des modèles statistiques en général, puis de les appliquer à un problème concret, tel que la séparation de la voix par rapport à la musique dans des enregistrements monophoniques de chansons. Apporter des solutions à ce problème, qui est assez difficile et peu étudié pour l'instant, peut être très utile pour faciliter l'analyse du contenu des chansons, par exemple dans le contexte de l'indexation audio. Les méthodes de séparation existantes donnent de bonnes performances à condition que les caractéristiques des modèles statistiques utilisés soient proches de celles des sources à séparer. Malheureusement, il n'est pas toujours possible de construire et d'utiliser en pratique de tels modèles, à cause de l'insuffisance des exemples d'apprentissage représentatifs et des ressources calculatoires. Pour remédier à ce problème, il est proposé dans cette thèse d'adapter a posteriori les modèles aux sources à séparer. Ainsi, un formalisme général d'adaptation est développé. En s'inspirant de techniques similaires utilisées en reconnaissance de la parole, ce formalisme est introduit sous la forme d'un critère d'adaptation Maximum A Posteriori (MAP). De plus, il est montré comment optimiser ce critère à l'aide de l'algorithme EM à différents niveaux de généralité. Ce formalisme d'adaptation est ensuite appliqué dans certaines formes particulières pour la séparation voix / musique. Les résultats obtenus montrent que pour cette tâche, l'utilisation des modèles adaptés permet d'augmenter significativement (au moins de 5 dB) les performances de séparation par rapport aux modèles non adaptés. Par ailleurs, il est observé que la séparation de la voix chantée facilite l'estimation de sa fréquence fondamentale (pitch), et que l'adaptation des modèles ne fait qu'améliorer ce résultat.

Robust speech recognition in acoustic environments that contain multiple speech sources and/or complex non-stationary noise is a difficult problem, but one of great practical interest. The formalism of probabilistic graphical models constitutes a relatively new and very powerful tool for better understanding and extending existing models, learning, and inference algorithms; and a bedrock for the creative, quasi-systematic development of new ones. In this thesis a collection of new graphical models and inference algorithms for robust speech recognition are presented. The problem of speech separation using multiple microphones is first treated. A family of variational algorithms for tractably combining multiple acoustic models of speech with observed sensor likelihoods is presented. The algorithms recover high quality estimates of the speech sources even when there are more sources than microphones, and have improved upon the state-of-the-art in terms of SNR gain by over 10 dB. Next the problem of background compensation in non-stationary acoustic environments is treated. A new dynamic noise adaptation (DNA) algorithm for robust noise compensation is presented, and shown to outperform several existing state-of-the-art front-end denoising systems on the new DNA + Aurora II and Aurora II-M extensions of the Aurora II task. Finally, the problem of speech recognition in speech using a single microphone is treated. The Iroquois system for multi-talker speech separation and recognition is presented. The system won the 2006 Pascal International Speech Separation Challenge, and amazingly, achieved super-human recognition performance on a majority of test cases in the task. The result marks a significant first in automatic speech recognition, and a milestone in computing.

Esta tese se concentra em aplicar os modelos de mistura de distribuições na segmentação de imagens SAR polarimétricas multi-look. Dentro deste contexto, utilizou-se o algoritmo SEM em conjunto com os estimadores obtidos pelo método dos momentos para calcular as estimativas dos parâmetros do modelo de mistura das distribuições Wishart, Kp ou G0p. Cada uma destas distribuições possui parâmetros específicos que as diferem no ajuste dos dados com graus de homogeneidade variados. A distribuição Wishart descreve bem regiões com características mais homogêneas, como cultivo. Esta distribuição é muito utilizada na análise de dados SAR polarimétricos multi-look. As distribuições Kp e G0p possuem um parâmetro de rugosidade que as permitem descrever tanto regiões mais heterogêneas, como vegetação e áreas urbanas, quanto regiões homogêneas. Além dos modelos de mistura de uma única família de distribuições, também foi analisado o caso de um dicionário contendo as três famílias. Há comparações do método SEM proposto para os diferentes modelos com os métodos da literatura k-médias e EM utilizando imagens reais da banda L. O método SEM com a mistura de distribuições G0p forneceu os melhores resultados quando os outliers da imagem são desconsiderados. A distribuição G0p foi a mais flexível ao ajuste dos diferentes tipos de alvo. A distribuição Wishart foi robusta às diferentes inicializações. O método k-médias com a distribuição Wishart é robusto à segmentação de imagens contendo outliers, mas não é muito flexível à variabilidade das regiões heterogêneas. O modelo de mistura do dicionário de famílias melhora a log-verossimilhança do método SEM, mas apresenta resultados parecidos com os do modelo de mistura G0p. Para todos os tipos de inicialização e grupos, a distribuição G0p predominou no processo de seleção das distribuições do dicionário de famílias. / The main focus of this thesis consists of the application of mixture models in multi-look polarimetric SAR image segmentation. Within this context, the SEM algorithm, together with the method of moments, were applied in the estimation of the Wishart, Kp and G0p mixture model parameters. Each one of these distributions has specific parameters that allows fitting data with different degrees of homogeneity. The Wishart distribution is suitable for modeling homogeneous regions, like crop fields for example. This distribution is widely used in multi-look polarimetric SAR data analysis. The distributions Kp and G0p have a roughness parameter that allows them to describe both heterogeneous regions, as vegetation and urban areas, and homogeneous regions. Besides adopting mixture models of a single family of distributions, the use of a dictionary with all the three family of distributions was proposed and analyzed. Also, a comparison between the performance of the proposed SEM method, considering the different models in real L-band images and two widely known techniques described in literature (k-means and EM algorithms), are shown and discussed. The proposed SEM method, considering a G0p mixture model combined with a outlier removal stage, provided the best classication results. The G0p distribution was the most flexible for fitting the different kinds of data. The Wishart distribution was robust for different initializations. The k-means algorithm with Wishart distribution is robust for segmentation of SAR images containing outliers, but it is not so flexible to variabilities in heterogeneous regions. The mixture model considering the dictionary of distributions improves the SEM method log-likelihood, but presents similar results to those of G0p mixture model. For all types of initializations and clusters, the G0p prevailed in the distribution selection process of the dictionary of distributions.