Adaptive Markov chain Monte Carlo and Bayesian filtering for state space models

This thesis is concerned with the state and parameter estimation in state space models. The estimation of states and parameters is an important task when mathematical modeling is applied to many different application areas such as the global positioning systems, target tracking, navigation, brain imaging, spread of infectious diseases, biological processes, telecommunications, audio signal processing, stochastic optimal control, machine learning, and physical systems. In Bayesian settings, the estimation of states or parameters amounts to computation of the posterior probability density function. Except for a very restricted number of models, it is impossible to compute this density function in a closed form. Hence, we need approximation methods. A state estimation problem involves estimating the states (latent variables) that are not directly observed in the output of the system. In this thesis, we use the Kalman filter, extended Kalman filter, Gauss–Hermite filters, and particle filters to estimate the states based on available measurements. Among these filters, particle filters are numerical methods for approximating the filtering distributions of non-linear non-Gaussian state space models via Monte Carlo. The performance of a particle filter heavily depends on the chosen importance distribution. For instance, inappropriate choice of the importance distribution can lead to the failure of convergence of the particle filter algorithm. In this thesis, we analyze the theoretical Lᵖ particle filter convergence with general importance distributions, where p ≥2 is an integer. A parameter estimation problem is considered with inferring the model parameters from measurements. For high-dimensional complex models, estimation of parameters can be done by Markov chain Monte Carlo (MCMC) methods. In its operation, the MCMC method requires the unnormalized posterior distribution of the parameters and a proposal distribution. In this thesis, we show how the posterior density function of the parameters of a state space model can be computed by filtering based methods, where the states are integrated out. This type of computation is then applied to estimate parameters of stochastic differential equations. Furthermore, we compute the partial derivatives of the log-posterior density function and use the hybrid Monte Carlo and scaled conjugate gradient methods to infer the parameters of stochastic differential equations. The computational efficiency of MCMC methods is highly depend on the chosen proposal distribution. A commonly used proposal distribution is Gaussian. In this kind of proposal, the covariance matrix must be well tuned. To tune it, adaptive MCMC methods can be used. In this thesis, we propose a new way of updating the covariance matrix using the variational Bayesian adaptive Kalman filter algorithm.

Statistical analysis of event-related potential elicited by verb-complement merge in Brazilian Portuguese

An interesting fact about language cognition is that stimulation involving incongruence in the merge operation between verb and complement has often been related to a negative event-related potential (ERP) of augmented amplitude and latency of ca. 400 ms - the N400. Using an automatic ERP latency and amplitude estimator to facilitate the recognition of waves with a low signal-to-noise ratio, the objective of the present study was to study the N400 statistically in 24 volunteers. Stimulation consisted of 80 experimental sentences (40 congruous and 40 incongruous), generated in Brazilian Portuguese, involving two distinct local verb-argument combinations (nominal object and pronominal object series). For each volunteer, the EEG was simultaneously acquired at 20 derivations, topographically localized according to the 10-20 International System. A computerized routine for automatic N400-peak marking (based on the ascendant zero-cross of the first waveform derivative) was applied to the estimated individual ERP waveform for congruous and incongruous sentences in both series for all ERP topographic derivations. Peak-to-peak N400 amplitude was significantly augmented (P < 0.05; one-sided Wilcoxon signed-rank test) due to incongruence in derivations F3, T3, C3, Cz, T5, P3, Pz, and P4 for nominal object series and in P3, Pz and P4 for pronominal object series. The results also indicated high inter-individual variability in ERP waveforms, suggesting that the usual procedure of grand averaging might not be considered a generally adequate approach. Hence, signal processing statistical techniques should be applied in neurolinguistic ERP studies allowing waveform analysis with low signal-to-noise ratio.

Effect of age on upper and lower eyelid saccades

To study the effect of age on the metrics of upper and lower eyelid saccades, eyelid movement of two groups of 30 subjects each were measured using computed image analysis. The patients were divided on the basis of age into a younger group (20-30 years) and an older group (60-91 years). Eyelid saccade functions were fitted by the damped harmonic oscillator model. Amplitude and peak velocity were used to compare the effect of age on the saccades of the upper and lower eyelid. There was no statistically significant difference in saccade amplitude between groups for the upper eyelid (mean ± SEM; upward, young = 9.18 ± 0.32 mm, older = 8.93 ± 0.31 mm, t = 0.56, P = 0.58; downward, young = 9.11 ± 0.27 mm, older = 8.86 ± 0.32 mm, t = 0.58, P = 0.56) However, there was a clear decline in the peak velocity of the upper eyelid saccades of older subjects (upward, young = 59.06 ± 2.34 mm/s, older = 50.12 ± 1.95 mm/s, t = 2.93, P = 0.005; downward, young = 71.78 ± 1.78 mm/s, older = 60.29 ± 2.62 mm/s, t = 3.63, P = 0.0006). In contrast, for the lower eyelid there was a clear increase of saccade amplitude in the elderly group (upward, young = 2.27 ± 0.09 mm, older = 2.98 ± 0.15 mm, t = 4.33, P < 0.0001; downward, young = 2.21 ± 0.10 mm, older = 2.96 ± 0.17 mm, t = 3.85, P < 0.001). These data suggest that the aging process affects the metrics of the lid saccades in a different manner according to the eyelid. In the upper eyelid the lower tension exerted by a weak aponeurosis is reflected only on the peak velocity of the saccades. In the lower eyelid, age is accompanied by an increase in saccade amplitude which indicates that the force transmission to the lid is not affected in the elderly.

Comparison of the Polar S810i monitor and the ECG for the analysis of heart rate variability in the time and frequency domains

The aim of the present study was to compare heart rate variability (HRV) at rest and during exercise using a temporal series obtained with the Polar S810i monitor and a signal from a LYNX® signal conditioner (BIO EMG 1000 model) with a channel configured for the acquisition of ECG signals. Fifteen healthy subjects aged 20.9 ± 1.4 years were analyzed. The subjects remained at rest for 20 min and performed exercise for another 20 min with the workload selected to achieve 60% of submaximal heart rate. RR series were obtained for each individual with a Polar S810i instrument and with an ECG analyzed with a biological signal conditioner. The HRV indices (rMSSD, pNN50, LFnu, HFnu, and LF/HF) were calculated after signal processing and analysis. The unpaired Student t-test and intraclass correlation coefficient were used for data analysis. No statistically significant differences were observed when comparing the values analyzed by means of the two devices for HRV at rest and during exercise. The intraclass correlation coefficient demonstrated satisfactory correlation between the values obtained by the devices at rest (pNN50 = 0.994; rMSSD = 0.995; LFnu = 0.978; HFnu = 0.978; LF/HF = 0.982) and during exercise (pNN50 = 0.869; rMSSD = 0.929; LFnu = 0.973; HFnu = 0.973; LF/HF = 0.942). The calculation of HRV values by means of temporal series obtained from the Polar S810i instrument appears to be as reliable as those obtained by processing the ECG signal captured with a signal conditioner.

Method for automatic detection of wheezing in lung sounds

The present report describes the development of a technique for automatic wheezing recognition in digitally recorded lung sounds. This method is based on the extraction and processing of spectral information from the respiratory cycle and the use of these data for user feedback and automatic recognition. The respiratory cycle is first pre-processed, in order to normalize its spectral information, and its spectrogram is then computed. After this procedure, the spectrogram image is processed by a two-dimensional convolution filter and a half-threshold in order to increase the contrast and isolate its highest amplitude components, respectively. Thus, in order to generate more compressed data to automatic recognition, the spectral projection from the processed spectrogram is computed and stored as an array. The higher magnitude values of the array and its respective spectral values are then located and used as inputs to a multi-layer perceptron artificial neural network, which results an automatic indication about the presence of wheezes. For validation of the methodology, lung sounds recorded from three different repositories were used. The results show that the proposed technique achieves 84.82% accuracy in the detection of wheezing for an isolated respiratory cycle and 92.86% accuracy for the detection of wheezes when detection is carried out using groups of respiratory cycles obtained from the same person. Also, the system presents the original recorded sound and the post-processed spectrogram image for the user to draw his own conclusions from the data.

Control of Bubbling Fluidized Bed Conditions

The objectives of this master’s thesis were to understand the importance of bubbling fluidized bed (BFB) conditions and to find out how digital image processing and acoustic emission technology can help in monitoring the bed quality. An acoustic emission (AE) measurement system and a bottom ash camera system were evaluated in acquiring information about the bed conditions. The theory part of the study describes the fundamentals of BFB boiler and evaluates the characteristics of bubbling bed. Causes and effects of bed material coarsening are explained. The ways and methods to monitor the behaviour of BFB are determined. The study introduces the operating principles of AE technology and digital image processing. The empirical part of the study describes an experimental arrangement and results of a case study at an industrial BFB boiler. Sand consumption of the boiler was reduced by optimization of bottom ash handling and sand feeding. Furthermore, data from the AE measurement system and the bottom ash camera system was collected. The feasibility of these two systems was evaluated. The particle size of bottom ash and the changes in particle size distribution were monitored during the test period. Neither of the systems evaluated was ready to serve in bed quality control accurately or fast enough. Particle size distributions according to the bottom ash camera did not correspond to the results of manual sieving. Comprehensive interpretation of the collected AE data requires much experience. Both technologies do have potential and with more research and development they may enable acquiring reliable and real-time information about the bed conditions. This information could help to maintain disturbance-free combustion process and to optimize bottom ash handling system.

Investigations in flow injection analysis (FIA): conventional FIA determination of chloride in water; sequential injection (SI)-FIA determination of mercury in water

Flow injection analysis (FIA) was applied to the determination of both chloride ion and mercury in water. Conventional FIA was employed for the chloride study. Investigations of the Fe3 +/Hg(SCN)2/CI-,450 nm spectrophotometric system for chloride determination led to the discovery of an absorbance in the 250-260 nm region when Hg(SCN)2 and CI- are combined in solution, in the absence of iron(III). Employing an in-house FIA system, absorbance observed at 254 nm exhibited a linear relation from essentially 0 - 2000 Jlg ml- 1 injected chloride. This linear range spanning three orders of magnitude is superior to the Fe3+/Hg(SCN)2/CI- system currently employed by laboratories worldwide. The detection limit obtainable with the proposed method was determin~d to be 0.16 Jlg ml- 1 and the relative standard deviation was determined to be 3.5 % over the concentration range of 0-200 Jig ml- 1. Other halogen ions were found to interfere with chloride determination at 254 nm whereas cations did not interfere. This system was successfully applied to the determination of chloride ion in laboratory water. Sequential injection (SI)-FIA was employed for mercury determination in water with the PSA Galahad mercury amalgamation, and Merlin mercury fluorescence detection systems. Initial mercury in air determinations involved injections of mercury saturated air directly into the Galahad whereas mercury in water determinations involved solution delivery via peristaltic pump to a gas/liquid separator, after reduction by stannous chloride. A series of changes were made to the internal hardware and valving systems of the Galahad mercury preconcentrator. Sequential injection solution delivery replaced the continuous peristaltic pump system and computer control was implemented to control and integrate all aspects of solution delivery, sample preconcentration and signal processing. Detection limits currently obtainable with this system are 0.1 ng ml-1 HgO.

A domain-general perspective on medial frontal brain activity during performance monitoring

Activity of the medial frontal cortex (MFC) has been implicated in attention regulation and performance monitoring. The MFC is thought to generate several event-related potential (ERPs) components, known as medial frontal negativities (MFNs), that are elicited when a behavioural response becomes difficult to control (e.g., following an error or shifting from a frequently executed response). The functional significance of MFNs has traditionally been interpreted in the context of the paradigm used to elicit a specific response, such as errors. In a series of studies, we consider the functional similarity of multiple MFC brain responses by designing novel performance monitoring tasks and exploiting advanced methods for electroencephalography (EEG) signal processing and robust estimation statistics for hypothesis testing. In study 1, we designed a response cueing task and used Independent Component Analysis (ICA) to show that the latent factors describing a MFN to stimuli that cued the potential need to inhibit a response on upcoming trials also accounted for medial frontal brain responses that occurred when individuals made a mistake or inhibited an incorrect response. It was also found that increases in theta occurred to each of these task events, and that the effects were evident at the group level and in single cases. In study 2, we replicated our method of classifying MFC activity to cues in our response task and showed again, using additional tasks, that error commission, response inhibition, and, to a lesser extent, the processing of performance feedback all elicited similar changes across MFNs and theta power. In the final study, we converted our response cueing paradigm into a saccade cueing task in order to examine the oscillatory dynamics of response preparation. We found that, compared to easy pro-saccades, successfully preparing a difficult anti-saccadic response was characterized by an increase in MFC theta and the suppression of posterior alpha power prior to executing the eye movement. These findings align with a large body of literature on performance monitoring and ERPs, and indicate that MFNs, along with their signature in theta power, reflects the general process of controlling attention and adapting behaviour without the need to induce error commission, the inhibition of responses, or the presentation of negative feedback.

Algorithmes de recommandation musicale

Ce mémoire est composé de trois articles qui s’unissent sous le thème de la recommandation musicale à grande échelle. Nous présentons d’abord une méthode pour effectuer des recommandations musicales en récoltant des étiquettes (tags) décrivant les items et en utilisant cette aura textuelle pour déterminer leur similarité. En plus d’effectuer des recommandations qui sont transparentes et personnalisables, notre méthode, basée sur le contenu, n’est pas victime des problèmes dont souffrent les systèmes de filtrage collaboratif, comme le problème du démarrage à froid (cold start problem). Nous présentons ensuite un algorithme d’apprentissage automatique qui applique des étiquettes à des chansons à partir d’attributs extraits de leur fichier audio. L’ensemble de données que nous utilisons est construit à partir d’une très grande quantité de données sociales provenant du site Last.fm. Nous présentons finalement un algorithme de génération automatique de liste d’écoute personnalisable qui apprend un espace de similarité musical à partir d’attributs audio extraits de chansons jouées dans des listes d’écoute de stations de radio commerciale. En plus d’utiliser cet espace de similarité, notre système prend aussi en compte un nuage d’étiquettes que l’utilisateur est en mesure de manipuler, ce qui lui permet de décrire de manière abstraite la sorte de musique qu’il désire écouter.

Les cartes fonctionnelles dans le cortex visuel du chat : nouvelles stratégies d’évaluation en imagerie optique et mise en évidence de l’organisation anatomo-fonctionnelle

Le regroupement des neurones de propriétés similaires est à l’origine de modules permettant d’optimiser l’analyse de l’information. La conséquence est la présence de cartes fonctionnelles dans le cortex visuel primaire de certains mammifères pour de nombreux paramètres tels que l’orientation, la direction du mouvement ou la position des stimuli (visuotopie). Le premier volet de cette thèse est consacré à caractériser l’organisation modulaire dans le cortex visuel primaire pour un paramètre fondamental, la suppression centre / pourtour et au delà du cortex visuel primaire (dans l’aire 21a), pour l’orientation et la direction. Toutes les études ont été effectuées à l’aide de l’imagerie optique des signaux intrinsèques sur le cortex visuel du chat anesthésié. La quantification de la modulation par la taille des stimuli à permis de révéler la présence de modules de forte et de faible suppression par le pourtour dans le cortex visuel primaire (aires 17 et 18). Ce type d’organisation n’avait été observé jusqu’ici que dans une aire de plus haut niveau hiérarchique chez le primate. Une organisation modulaire pour l’orientation, similaire à celle observée dans le cortex visuel primaire a été révélée dans l’aire 21a. Par contre, contrairement à l’aire 18, l’aire 21a ne semblait pas être organisée en domaine de direction. L’ensemble de ces résultats pourront permettre d’alimenter les connaissances sur l’organisation anatomo-fonctionnelle du cortex visuel du chat mais également de mieux comprendre les facteurs qui déterminent la présence d’une organisation modulaire. Le deuxième volet abordé dans cette thèse s’est intéressé à l’amélioration de l’aspect quantitatif apporté par l’analyse temporelle en imagerie optique des signaux intrinsèques. Cette nouvelle approche, basée sur l’analyse de Fourier a permis d’augmenter considérablement le rapport signal / bruit des enregistrements. Toutefois, cette analyse ne s’est basée jusqu’ici que sur la quantification d’une seule harmonique ce qui a limité son emploi à la cartographie de l’orientation et de rétinotopie uniquement. En exploitant les plus hautes harmoniques, un modèle a été proposé afin d’estimer la taille des champs récepteurs et la sélectivité à la direction. Ce modèle a par la suite été validé par des approches conventionnelles dans le cortex visuel primaire.

Génération et reconnaissance de rythmes au moyen de réseaux de neurones à réservoir

Les fichiers sons qui accompagne mon document sont au format midi. Le programme que nous avons développés pour ce travail est en language Python.

Apprentissage statistique pour l'étiquetage de musique et la recommandation

Mémoire numérisé par la Division de la gestion de documents et des archives de l'Université de Montréal

Apprentissage de représentations musicales à l'aide d'architectures profondes et multiéchelles

L'apprentissage machine (AM) est un outil important dans le domaine de la recherche d'information musicale (Music Information Retrieval ou MIR). De nombreuses tâches de MIR peuvent être résolues en entraînant un classifieur sur un ensemble de caractéristiques. Pour les tâches de MIR se basant sur l'audio musical, il est possible d'extraire de l'audio les caractéristiques pertinentes à l'aide de méthodes traitement de signal. Toutefois, certains aspects musicaux sont difficiles à extraire à l'aide de simples heuristiques. Afin d'obtenir des caractéristiques plus riches, il est possible d'utiliser l'AM pour apprendre une représentation musicale à partir de l'audio. Ces caractéristiques apprises permettent souvent d'améliorer la performance sur une tâche de MIR donnée. Afin d'apprendre des représentations musicales intéressantes, il est important de considérer les aspects particuliers à l'audio musical dans la conception des modèles d'apprentissage. Vu la structure temporelle et spectrale de l'audio musical, les représentations profondes et multiéchelles sont particulièrement bien conçues pour représenter la musique. Cette thèse porte sur l'apprentissage de représentations de l'audio musical. Des modèles profonds et multiéchelles améliorant l'état de l'art pour des tâches telles que la reconnaissance d'instrument, la reconnaissance de genre et l'étiquetage automatique y sont présentés.

Amorçage cognitif pour amélioration de l’acquisition de la connaissance dans un système tutoriel intelligent

Cette thèse vise à définir une nouvelle méthode d’enseignement pour les systèmes tutoriels intelligents dans le but d’améliorer l’acquisition des connaissances. L’apprentissage est un phénomène complexe faisant intervenir des mécanismes émotionnels et cognitifs de nature consciente et inconsciente. Nous nous intéressons à mieux comprendre les mécanismes inconscients du raisonnement lors de l’acquisition des connaissances. L’importance de ces processus inconscients pour le raisonnement est bien documentée en neurosciences, mais demeure encore largement inexplorée dans notre domaine de recherche. Dans cette thèse, nous proposons la mise en place d’une nouvelle approche pédagogique dans le domaine de l’éducation implémentant une taxonomie neuroscientifique de la perception humaine. Nous montrons que cette nouvelle approche agit sur le raisonnement et, à tour de rôle, améliore l’apprentissage général et l’induction de la connaissance dans un environnement de résolution de problème. Dans une première partie, nous présentons l’implémentation de notre nouvelle méthode dans un système tutoriel visant à améliorer le raisonnement pour un meilleur apprentissage. De plus, compte tenu de l’importance des mécanismes émotionnels dans l’apprentissage, nous avons également procédé dans cette partie à la mesure des émotions par des capteurs physiologiques. L’efficacité de notre méthode pour l’apprentissage et son impact positif observé sur les émotions a été validée sur trente et un participants. Dans une seconde partie, nous allons plus loin dans notre recherche en adaptant notre méthode visant à améliorer le raisonnement pour une meilleure induction de la connaissance. L’induction est un type de raisonnement qui permet de construire des règles générales à partir d’exemples spécifiques ou de faits particuliers. Afin de mieux comprendre l’impact de notre méthode sur les processus cognitifs impliqués dans ce type de raisonnement, nous avons eu recours à des capteurs cérébraux pour mesurer l’activité du cerveau des utilisateurs. La validation de notre approche réalisée sur quarante-trois volontaires montre l’efficacité de notre méthode pour l’induction de la connaissance et la viabilité de mesurer le raisonnement par des mesures cérébrales suite à l’application appropriée d’algorithmes de traitement de signal. Suite à ces deux parties, nous clorons la thèse par une discussion applicative en décrivant la mise en place d’un nouveau système tutoriel intelligent intégrant les résultats de nos travaux.

The Wolf and the Whale: aesthetic relationships between electroacoustic music and poetry inspired by the Canadian landscape

La version intégrale de cette thèse est disponible uniquement pour consultation individuelle à la Bibliothèque de musique de l’Université de Montréal (www.bib.umontreal.ca/MU).