Modeling High-Dimensional Audio Sequences with Recurrent Neural Networks

Cette thèse étudie des modèles de séquences de haute dimension basés sur des réseaux de neurones récurrents (RNN) et leur application à la musique et à la parole. Bien qu'en principe les RNN puissent représenter les dépendances à long terme et la dynamique temporelle complexe propres aux séquences d'intérêt comme la vidéo, l'audio et la langue naturelle, ceux-ci n'ont pas été utilisés à leur plein potentiel depuis leur introduction par Rumelhart et al. (1986a) en raison de la difficulté de les entraîner efficacement par descente de gradient. Récemment, l'application fructueuse de l'optimisation Hessian-free et d'autres techniques d'entraînement avancées ont entraîné la recrudescence de leur utilisation dans plusieurs systèmes de l'état de l'art. Le travail de cette thèse prend part à ce développement. L'idée centrale consiste à exploiter la flexibilité des RNN pour apprendre une description probabiliste de séquences de symboles, c'est-à-dire une information de haut niveau associée aux signaux observés, qui en retour pourra servir d'à priori pour améliorer la précision de la recherche d'information. Par exemple, en modélisant l'évolution de groupes de notes dans la musique polyphonique, d'accords dans une progression harmonique, de phonèmes dans un énoncé oral ou encore de sources individuelles dans un mélange audio, nous pouvons améliorer significativement les méthodes de transcription polyphonique, de reconnaissance d'accords, de reconnaissance de la parole et de séparation de sources audio respectivement. L'application pratique de nos modèles à ces tâches est détaillée dans les quatre derniers articles présentés dans cette thèse. Dans le premier article, nous remplaçons la couche de sortie d'un RNN par des machines de Boltzmann restreintes conditionnelles pour décrire des distributions de sortie multimodales beaucoup plus riches. Dans le deuxième article, nous évaluons et proposons des méthodes avancées pour entraîner les RNN. Dans les quatre derniers articles, nous examinons différentes façons de combiner nos modèles symboliques à des réseaux profonds et à la factorisation matricielle non-négative, notamment par des produits d'experts, des architectures entrée/sortie et des cadres génératifs généralisant les modèles de Markov cachés. Nous proposons et analysons également des méthodes d'inférence efficaces pour ces modèles, telles la recherche vorace chronologique, la recherche en faisceau à haute dimension, la recherche en faisceau élagué et la descente de gradient. Finalement, nous abordons les questions de l'étiquette biaisée, du maître imposant, du lissage temporel, de la régularisation et du pré-entraînement.

Développement d'une nouvelle méthode de caractérisation tissulaire basée sur l'élastographie ultrasonore : application pour le dépistage précoce du cancer du sein

Le cancer du sein est le cancer le plus fréquent chez la femme. Il demeure la cause de mortalité la plus importante chez les femmes âgées entre 35 et 55 ans. Au Canada, plus de 20 000 nouveaux cas sont diagnostiqués chaque année. Les études scientifiques démontrent que l'espérance de vie est étroitement liée à la précocité du diagnostic. Les moyens de diagnostic actuels comme la mammographie, l'échographie et la biopsie comportent certaines limitations. Par exemple, la mammographie permet de diagnostiquer la présence d’une masse suspecte dans le sein, mais ne peut en déterminer la nature (bénigne ou maligne). Les techniques d’imagerie complémentaires comme l'échographie ou l'imagerie par résonance magnétique (IRM) sont alors utilisées en complément, mais elles sont limitées quant à la sensibilité et la spécificité de leur diagnostic, principalement chez les jeunes femmes (< 50 ans) ou celles ayant un parenchyme dense. Par conséquent, nombreuses sont celles qui doivent subir une biopsie alors que leur lésions sont bénignes. Quelques voies de recherche sont privilégiées depuis peu pour réduire l`incertitude du diagnostic par imagerie ultrasonore. Dans ce contexte, l’élastographie dynamique est prometteuse. Cette technique est inspirée du geste médical de palpation et est basée sur la détermination de la rigidité des tissus, sachant que les lésions en général sont plus rigides que le tissu sain environnant. Le principe de cette technique est de générer des ondes de cisaillement et d'en étudier la propagation de ces ondes afin de remonter aux propriétés mécaniques du milieu via un problème inverse préétabli. Cette thèse vise le développement d'une nouvelle méthode d'élastographie dynamique pour le dépistage précoce des lésions mammaires. L'un des principaux problèmes des techniques d'élastographie dynamiques en utilisant la force de radiation est la forte atténuation des ondes de cisaillement. Après quelques longueurs d'onde de propagation, les amplitudes de déplacement diminuent considérablement et leur suivi devient difficile voir impossible. Ce problème affecte grandement la caractérisation des tissus biologiques. En outre, ces techniques ne donnent que l'information sur l'élasticité tandis que des études récentes montrent que certaines lésions bénignes ont les mêmes élasticités que des lésions malignes ce qui affecte la spécificité de ces techniques et motive la quantification de d'autres paramètres mécaniques (e.g.la viscosité). Le premier objectif de cette thèse consiste à optimiser la pression de radiation acoustique afin de rehausser l'amplitude des déplacements générés. Pour ce faire, un modèle analytique de prédiction de la fréquence de génération de la force de radiation a été développé. Une fois validé in vitro, ce modèle a servi pour la prédiction des fréquences optimales pour la génération de la force de radiation dans d'autres expérimentations in vitro et ex vivo sur des échantillons de tissu mammaire obtenus après mastectomie totale. Dans la continuité de ces travaux, un prototype de sonde ultrasonore conçu pour la génération d'un type spécifique d'ondes de cisaillement appelé ''onde de torsion'' a été développé. Le but est d'utiliser la force de radiation optimisée afin de générer des ondes de cisaillement adaptatives, et de monter leur utilité dans l'amélioration de l'amplitude des déplacements. Contrairement aux techniques élastographiques classiques, ce prototype permet la génération des ondes de cisaillement selon des parcours adaptatifs (e.g. circulaire, elliptique,…etc.) dépendamment de la forme de la lésion. L’optimisation des dépôts énergétiques induit une meilleure réponse mécanique du tissu et améliore le rapport signal sur bruit pour une meilleure quantification des paramètres viscoélastiques. Il est aussi question de consolider davantage les travaux de recherches antérieurs par un appui expérimental, et de prouver que ce type particulier d'onde de torsion peut mettre en résonance des structures. Ce phénomène de résonance des structures permet de rehausser davantage le contraste de déplacement entre les masses suspectes et le milieu environnant pour une meilleure détection. Enfin, dans le cadre de la quantification des paramètres viscoélastiques des tissus, la dernière étape consiste à développer un modèle inverse basé sur la propagation des ondes de cisaillement adaptatives pour l'estimation des paramètres viscoélastiques. L'estimation des paramètres viscoélastiques se fait via la résolution d'un problème inverse intégré dans un modèle numérique éléments finis. La robustesse de ce modèle a été étudiée afin de déterminer ces limites d'utilisation. Les résultats obtenus par ce modèle sont comparés à d'autres résultats (mêmes échantillons) obtenus par des méthodes de référence (e.g. Rheospectris) afin d'estimer la précision de la méthode développée. La quantification des paramètres mécaniques des lésions permet d'améliorer la sensibilité et la spécificité du diagnostic. La caractérisation tissulaire permet aussi une meilleure identification du type de lésion (malin ou bénin) ainsi que son évolution. Cette technique aide grandement les cliniciens dans le choix et la planification d'une prise en charge adaptée.

Fish Driven Bioturbation as a Viable Aquaculture Technology for Increasing Production and Improving Ecological Sustainability of Giant Tiger Shrimp Penaeus Monodon (FABRlClUS)

A general introduction to the problems faced in the shrimp culture due to waste formation and its consequent environmental hazards and production problems of Giant tiger shrimp, Penaeus monodon is highlighted by the author in this thesis. The objective of the present work was to assess the potential of brackish water finfish to improve bottom soil conditions and thereby increase the growth and production of Penaeus monodon. The salient findings of the present study are summarized in chapter 7. This is followed by the references cited in the thesis and list ofpublications originated from the present study.

Development of Hierarchical Clustering Techniques for Gridded Data from Mixed Data Sequences

Knowledge discovery in databases is the non-trivial process of identifying valid, novel potentially useful and ultimately understandable patterns from data. The term Data mining refers to the process which does the exploratory analysis on the data and builds some model on the data. To infer patterns from data, data mining involves different approaches like association rule mining, classification techniques or clustering techniques. Among the many data mining techniques, clustering plays a major role, since it helps to group the related data for assessing properties and drawing conclusions. Most of the clustering algorithms act on a dataset with uniform format, since the similarity or dissimilarity between the data points is a significant factor in finding out the clusters. If a dataset consists of mixed attributes, i.e. a combination of numerical and categorical variables, a preferred approach is to convert different formats into a uniform format. The research study explores the various techniques to convert the mixed data sets to a numerical equivalent, so as to make it equipped for applying the statistical and similar algorithms. The results of clustering mixed category data after conversion to numeric data type have been demonstrated using a crime data set. The thesis also proposes an extension to the well known algorithm for handling mixed data types, to deal with data sets having only categorical data. The proposed conversion has been validated on a data set corresponding to breast cancer. Moreover, another issue with the clustering process is the visualization of output. Different geometric techniques like scatter plot, or projection plots are available, but none of the techniques display the result projecting the whole database but rather demonstrate attribute-pair wise analysis

Analysis of Stochastic Volatility Sequences Generated by Product Autoregressive Models

The classical methods of analysing time series by Box-Jenkins approach assume that the observed series uctuates around changing levels with constant variance. That is, the time series is assumed to be of homoscedastic nature. However, the nancial time series exhibits the presence of heteroscedasticity in the sense that, it possesses non-constant conditional variance given the past observations. So, the analysis of nancial time series, requires the modelling of such variances, which may depend on some time dependent factors or its own past values. This lead to introduction of several classes of models to study the behaviour of nancial time series. See Taylor (1986), Tsay (2005), Rachev et al. (2007). The class of models, used to describe the evolution of conditional variances is referred to as stochastic volatility modelsThe stochastic models available to analyse the conditional variances, are based on either normal or log-normal distributions. One of the objectives of the present study is to explore the possibility of employing some non-Gaussian distributions to model the volatility sequences and then study the behaviour of the resulting return series. This lead us to work on the related problem of statistical inference, which is the main contribution of the thesis

A generic polynomial solution for the differential equation of hypergeometric type and six sequences of orthogonal polynomials related to it

In this work, we present a generic formula for the polynomial solution families of the well-known differential equation of hypergeometric type s(x)y"n(x) + t(x)y'n(x) - lnyn(x) = 0 and show that all the three classical orthogonal polynomial families as well as three finite orthogonal polynomial families, extracted from this equation, can be identified as special cases of this derived polynomial sequence. Some general properties of this sequence are also given.

Mechanisms of residue mulch-induced cereal growth increases in West Africa

The use of crop residues (CR) has been widely reported as a means of increasing crop yields across West Africa. However, little has been done to compare the magnitude and mechanisms of CR effects systematically in the different agro-ecological zones of the region. To this end, a series of field trials with millet (Pennisetum glaucum L.), sorghum [Sorghum bicolor (L.) Moench], and maize (Zea mays L.) was conducted over a 4-yr period in the Sahelian, Sudanian, and Guinean zones of West Africa. Soils ranged in pH from 4.1 to 5.4 along a rainfall gradient from 510 to 1300 mm. Treatments in the factorial experiments were three CR rates (0,500, and 2000 kg ha^-1)and several levels of phosphorus and nitrogen. The results showed CR-induced total dry matter (TDM) increases in cereals up to 73% for the Sahel compared with a maximum of 16% in the wetter Sudanian and Guinean zones. Residue effects on weakly buffered Sahelian soils were due to improved P availability and to a protection of seedlings against wind erosion. Additional effects of CR mulching on topsoil properties in the Sahel were a decrease in peak temperatures by 4°C and increased water availability. These mulch effects on soil chemical and physical properties strongly decreased from North to South. Likely explanations for this decrease are the decline of dust deposition and wind erosion hazards, the higher soil clay content, lower air temperature, and a faster decomposition rate of mulch material with increasing rainfall from the Sahel to the Sudanian and Guinean zones.

Causes of legume-rotation effects in increasing cereal yields across the Sudanian, Sahelian and Guinean zone of West Africa

On-farm experiments and pot trials were conducted on eight West African soils to explore the mechanisms governing the often reported legume rotation-induced cereal growth increases in this region. Crops comprised pearl millet (Pennisetum glaucum L.), sorghum (Sorghum bicolor Moench), maize (Zea mays L.), cowpea (Vigna unguiculata Walp.) and groundnut (Arachis hypogaea L.). In groundnut trials the observed 26 to 85% increases in total dry matter (TDM) of rotation cereals (RC) compared with continuous cereals (CC) in the 4th year appeared to be triggered by site- and crop-specific early season differences in nematode infestation (up to 6-fold lower in RC than in CC), enhanced Nmin and a 7% increase in mycorrhizal (AM) infection. In cowpea trials yield effects on millet and differences in nematode numbers, Nmin and AM were much smaller. Rhizosphere studies indicated effects on pH and acid phosphatase activity as secondary causes for the observed growth differences between RC and CC. In the study region legume-rotation effects on cereals seemed to depend on the capability of the legume to suppress nematodes and to enhance early N and P availability for the subsequent cereal.

Increasing scepticism toward potential liars: effects of existential threat on veracity judgments and the moderating role of honesty norm activation

With the present research, we investigated effects of existential threat on veracity judgments. According to several meta-analyses, people judge potentially deceptive messages of other people as true rather than as false (so-called truth bias). This judgmental bias has been shown to depend on how people weigh the error of judging a true message as a lie (error 1) and the error of judging a lie as a true message (error 2). The weight of these errors has been further shown to be affected by situational variables. Given that research on terror management theory has found evidence that mortality salience (MS) increases the sensitivity toward the compliance of cultural norms, especially when they are of focal attention, we assumed that when the honesty norm is activated, MS affects judgmental error weighing and, consequently, judgmental biases. Specifically, activating the norm of honesty should decrease the weight of error 1 (the error of judging a true message as a lie) and increase the weight of error 2 (the error of judging a lie as a true message) when mortality is salient. In a first study, we found initial evidence for this assumption. Furthermore, the change in error weighing should reduce the truth bias, automatically resulting in better detection accuracy of actual lies and worse accuracy of actual true statements. In two further studies, we manipulated MS and honesty norm activation before participants judged several videos containing actual truths or lies. Results revealed evidence for our prediction. Moreover, in Study 3, the truth bias was increased after MS when group solidarity was previously emphasized.

Online Learning of Non-stationary Sequences

We consider an online learning scenario in which the learner can make predictions on the basis of a fixed set of experts. The performance of each expert may change over time in a manner unknown to the learner. We formulate a class of universal learning algorithms for this problem by expressing them as simple Bayesian algorithms operating on models analogous to Hidden Markov Models (HMMs). We derive a new performance bound for such algorithms which is considerably simpler than existing bounds. The bound provides the basis for learning the rate at which the identity of the optimal expert switches over time. We find an analytic expression for the a priori resolution at which we need to learn the rate parameter. We extend our scalar switching-rate result to models of the switching-rate that are governed by a matrix of parameters, i.e. arbitrary homogeneous HMMs. We apply and examine our algorithm in the context of the problem of energy management in wireless networks. We analyze the new results in the framework of Information Theory.

People Recognition in Image Sequences by Supervised Learning

We describe a system that learns from examples to recognize people in images taken indoors. Images of people are represented by color-based and shape-based features. Recognition is carried out through combinations of Support Vector Machine classifiers (SVMs). Different types of multiclass strategies based on SVMs are explored and compared to k-Nearest Neighbors classifiers (kNNs). The system works in real time and shows high performance rates for people recognition throughout one day.

Alternative ways to estimate change points in multinomial sequences. An application to an authorship attribution problem

The statistical analysis of literary style is the part of stylometry that compares measurable characteristics in a text that are rarely controlled by the author, with those in other texts. When the goal is to settle authorship questions, these characteristics should relate to the author’s style and not to the genre, epoch or editor, and they should be such that their variation between authors is larger than the variation within comparable texts from the same author. For an overview of the literature on stylometry and some of the techniques involved, see for example Mosteller and Wallace (1964, 82), Herdan (1964), Morton (1978), Holmes (1985), Oakes (1998) or Lebart, Salem and Berry (1998). Tirant lo Blanc, a chivalry book, is the main work in catalan literature and it was hailed to be “the best book of its kind in the world” by Cervantes in Don Quixote. Considered by writters like Vargas Llosa or Damaso Alonso to be the first modern novel in Europe, it has been translated several times into Spanish, Italian and French, with modern English translations by Rosenthal (1996) and La Fontaine (1993). The main body of this book was written between 1460 and 1465, but it was not printed until 1490. There is an intense and long lasting debate around its authorship sprouting from its first edition, where its introduction states that the whole book is the work of Martorell (1413?-1468), while at the end it is stated that the last one fourth of the book is by Galba (?-1490), after the death of Martorell. Some of the authors that support the theory of single authorship are Riquer (1990), Chiner (1993) and Badia (1993), while some of those supporting the double authorship are Riquer (1947), Coromines (1956) and Ferrando (1995). For an overview of this debate, see Riquer (1990). Neither of the two candidate authors left any text comparable to the one under study, and therefore discriminant analysis can not be used to help classify chapters by author. By using sample texts encompassing about ten percent of the book, and looking at word length and at the use of 44 conjunctions, prepositions and articles, Ginebra and Cabos (1998) detect heterogeneities that might indicate the existence of two authors. By analyzing the diversity of the vocabulary, Riba and Ginebra (2000) estimates that stylistic boundary to be near chapter 383. Following the lead of the extensive literature, this paper looks into word length, the use of the most frequent words and into the use of vowels in each chapter of the book. Given that the features selected are categorical, that leads to three contingency tables of ordered rows and therefore to three sequences of multinomial observations. Section 2 explores these sequences graphically, observing a clear shift in their distribution. Section 3 describes the problem of the estimation of a suden change-point in those sequences, in the following sections we propose various ways to estimate change-points in multinomial sequences; the method in section 4 involves fitting models for polytomous data, the one in Section 5 fits gamma models onto the sequence of Chi-square distances between each row profiles and the average profile, the one in Section 6 fits models onto the sequence of values taken by the first component of the correspondence analysis as well as onto sequences of other summary measures like the average word length. In Section 7 we fit models onto the marginal binomial sequences to identify the features that distinguish the chapters before and after that boundary. Most methods rely heavily on the use of generalized linear models

Towards Detecting Changes in Underwater Image Sequences

Detecting changes between images of the same scene taken at different times is of great interest for monitoring and understanding the environment. It is widely used for on-land application but suffers from different constraints. Unfortunately, Change detection algorithms require highly accurate geometric and photometric registration. This requirement has precluded their use in underwater imagery in the past. In this paper, the change detection techniques available nowadays for on-land application were analyzed and a method to automatically detect the changes in sequences of underwater images is proposed. Target application scenarios are habitat restoration sites, or area monitoring after sudden impacts from hurricanes or ship groundings. The method is based on the creation of a 3D terrain model from one image sequence over an area of interest. This model allows for synthesizing textured views that correspond to the same viewpoints of a second image sequence. The generated views are photometrically matched and corrected against the corresponding frames from the second sequence. Standard change detection techniques are then applied to find areas of difference. Additionally, the paper shows that it is possible to detect false positives, resulting from non-rigid objects, by applying the same change detection method to the first sequence exclusively. The developed method was able to correctly find the changes between two challenging sequences of images from a coral reef taken one year apart and acquired with two different cameras

Online Robust 3D Mapping Using Structure from Motion Cues

This paper presents a complete solution for creating accurate 3D textured models from monocular video sequences. The methods are developed within the framework of sequential structure from motion, where a 3D model of the environment is maintained and updated as new visual information becomes available. The camera position is recovered by directly associating the 3D scene model with local image observations. Compared to standard structure from motion techniques, this approach decreases the error accumulation while increasing the robustness to scene occlusions and feature association failures. The obtained 3D information is used to generate high quality, composite visual maps of the scene (mosaics). The visual maps are used to create texture-mapped, realistic views of the scene