The de-facto standard for dysarthric intelligibility assessment is a subjective intelligibility test, performed by an expert. Subjective tests are often costly, biased and inconsistent because of their perceptual nature. Automatic objective assessment methods, in contrast, are repeatable and relatively cheap. Objective methods can be broken down into two subcategories: reference-free, and reference based. Reference-free methods employ estimation procedures that do not require information about the target speech material. This potentially makes the problem more difficult, and consequently, there is a deficit of research into reference-free dysarthric intelligibility estimation. In this thesis, we focus on the reference-free intelligibility estimation approach. To make the problem more tractable, we focus on the dysarthrias of cerebral palsy (CP). First, a popular standard for blind speech quality estimation, the ITU-T P.563 standard, is examined for possible application to dysarthric intelligibility estimation. The internal structure of the standard is discussed, along with the relevance of its internal features to intelligibility estimation. Afterwards, several novel features expected to relate to some of the acoustic properties of dysarthric speech are proposed. Proposed features are based on the high-order statistics of parameters derived from linear prediction (LP) analysis, and a mel-frequency filterbank. In order to gauge the complimentariness of P.563 and proposed features, a linear intelligibility model is proposed and tested. Intelligibility is expressed as a linear combination of acoustic features, which are selected from a feature pool using speaker-dependent and speaker-independent validation methods. An intelligibility estimator constructed with only P.563 features serves as the `baseline'. When proposed features are added to the feature pool, performance is shown to improve substantially for both speaker-dependent and speaker-independent methods when compared to the baseline. Results are also shown to compare favourably with those reported in the literature. / Thesis (Master, Electrical & Computer Engineering) -- Queen's University, 2011-09-28 18:44:51.103

Thesis advisor: Gabor T. Marth / To help improve the analysis of forward genetic screens, we have developed an efficient and automated pipeline for mutational profiling using our reference guided tools including MOSAIK and FREEBAYES. Studies using next generation sequencing technologies currently employ either reference guided alignment or de novo assembly to analyze the massive amount of short read data produced by second generation sequencing technologies; the far more common approach being reference guided alignment due to the massive computational and sequencing costs associated with de novo assembly. The success of reference guided alignment is dependent on three factors; the accuracy of the reference, the ability of the mapper to correctly place a read, and the degree to which a variant allele differs from the reference. Reference assemblies are not perfect and none are entirely complete. Moreover, read mappers can only map reads in genomic locations that are unique enough to confidently place reads; paralogous sections, such as related gene families, cannot be characterized and are often ignored. Further, variant alleles that drastically alter the subject's DNA, such as insertions or deletions (INDELs), will not map to the reference and are either entirely missed or require further downstream analysis to characterize. Most importantly, reference guided methods are restricted to organisms for which such reference genomes have been assembled. The current alternative, de novo assembly of a genome, is prohibitively expensive for most labs requiring deep read coverage from numerous different library preparations as well as massive computing power. To address the shortcomings of current methods, while eliminating the costs intrinsic to de novo sequence assembly, we developed RUFUS, a novel, completely reference-independent variant discovery tool. RUFUS directly compares raw sequence data from two or more samples and identifies groups of reads unique to one or the other sample. RUFUS has at least the same variant detection sensitivity as mapping methods, with greatly increased specificity for SNPs and INDEL variation events. RUFUS is also capable of extremely sensitive copy number detection, without any restriction on event length. By modeling the underlying k-mer distribution, RUFUS produces a specific copy number spectrum for each individual sample. Applying a Bayesian detection method to detect changes in k-mer content between two samples, RUFUS produces copy number calls that are equally as sensitive as traditional copy number detection methods with far fewer false positives. Our data suggest that RUFUS' reference-free approach to variant discovery is able to substantially improve upon existing variant detection methods: reducing reference biases, reducing false positive variants, and detecting copy number variants with excellent sensitivity and specificity. / Thesis (PhD) — Boston College, 2014. / Submitted to: Boston College. Graduate School of Arts and Sciences. / Discipline: Biology.

In high-diversity populations, a complete accounting of de novo mutations can be difficult to obtain. Most analyses involve identifying such mutations by sequencing pedigrees on second-generation sequencing platforms and aligning the short reads to a reference assembly, the genomic sequence of a canonical member (or members) of a species. Often, large regions of the genomes under study may be greatly diverged from the reference sequence, or not represented at all (e.g. the HLA, antigenic genes, or other regions under balancing selective pressure). If the haplotypic background upon which a mutation occurs is absent, events can easily be missed (as reads have nowhere to align) and false-positives may abound (as the software forces the reads to align elsewhere). This thesis presents a novel method for de novo mutation discovery and allele identification. Rather than relying on alignment, our method is based on the de novo assembly of short-read sequence data using a multi-color de Bruijn graph. In this data structure, each sample is assigned a unique index (or "color"), reads from each sample are decomposed into smaller subsequences of length k (or "kmers"), and color-specific adjacency information between kmers is recorded. Mutations can be discovered in the graph itself by searching for characteristic motifs (e.g. a "bubble motifs", indicative of a SNP or indel, and "linear motifs" indicative of allelic and non-allelic recombination). De novo mutations differ from inherited mutations in that the kmers spanning the variant allele are absent in the parents; in a sense, they facilitate their own discovery by generating "novel" sequence. We exploit this fact to limit processing of the graph to only those regions containing these novel kmers. We verified our approach using simulations, validation, and visualization. On the simulations, we developed genome and read generation software driven by empirical distributions computed from real data to emit genomes with realistic features: recombinations, de novo variants, read fragment sizes, sequencing errors, and coverage profiles. In 20 artifical samples, we determined our sensitivity and specificity for novel kmer recovery to be approximately 98% and 100% at worst, respectively. Not every novel stretch can be reconstituted as a variant, owing to errors and homology in the graph. In simulations, our false discovery rate was 10% for "bubble" events and 12% for "linear" events. On validation, we obtained a high-quality draft assembly for a single P. falciparum child using a third-generation sequencing platform. We discovered three de novo events in the draft assembly, all three of which are recapitulated in our calls on the second-generation sequencing data for the same sample; no false-positives are present. On visualization, we developed an interactive web application capable of rendering a multi-color subgraph that assists in visually distinguishing between true variation and sequencing artifacts. We applied our caller to real datasets: 115 progeny across four previously analyzed experimental crosses of Plasmodium falciparum. We demonstrate our ability to access subtelomeric compartments of the genome, regions harboring antigenic genes under tremendous selective pressure, thus highly divergent between geographically distinct isolates and routinely masked and ignored in reference-based analyses. We also show our caller's ability to recover an important form of structural de novo variation: non-allelic homologous recombination (NAHR) events, an important mechanism for the pathogen to diversify its own antigenic repertoire. We demonstrate our ability to recover the few events in these samples known to exist, and overturn some previous findings indicating exchanges between "core" (non-subtelomeric) genes. We compute the SNP mutation rate to be approximately 2.91 per sample, insertion and deletion mutation rates to be 0.55 and 1.04 per sample, respectively, multi-nucleotide polymorphisms to be 0.72 per sample, and NAHR events to be 0.33 per sample. These findings are consistent across crosses. Finally, we investigated our method's scaling capabilities by processing a quintet of previously analyzed Pan troglodytes verus (western chimpanzee) samples. The genome of the chimpanzee is two orders of magnitude larger than the malaria parasite's (3, 300 Mbp versus 23 Mbp), diploid rather than haploid, poorly assembled, and the read dataset is lower coverage (20x versus 120x). Comparing to Sequenom validation data as well as visual validation, our sensitivity is expectedly low. However, this can be attributed to overaggressiveness in data cleaning applied by the de novo assembler atop which our software is built. We discuss the precise changes that would likely need to be made in future work to adapt our method to low-coverage samples.

Le sous-titrage d’images est la tâche de l’intelligence artificielle (IA) qui consiste à décrire des images en langage naturel. Cette tâche d’IA a plusieurs applications sociétales utiles, telles que l’accessibilité pour les malvoyants, la génération automatisée de contenu, l’interaction humain-robot et l’analyse d’imagerie médicale. Au cours des huit dernières années, la recherche sur le sous-titrage d'images a connu d'énormes progrès dans la création de modèles solides, la collecte d'ensembles de données à grande échelle ainsi que le développement de mesures d'évaluation automatique. Malgré ces progrès remarquables, la recherche sur le sous-titrage d'images est confrontée à deux défis majeurs: 1) Comment construire des modèles efficaces en termes de paramètres, et 2) Comment construire des métriques d'évaluation automatique robustes. Dans cette thèse, nous apportons notre contribution à la résolution de chacun de ces défis. Premièrement, nous proposons une méthode efficace en termes de paramètres (MAPL \cite{mapl}) qui adapte des modèles pré-entraînés unimodaux de vision uniquement et de langage uniquement pour la tâche multimodale de sous-titrage d'images. MAPL apprend un mappage léger entre les espaces de représentation des modèles unimodaux. Ainsi, MAPL peut exploiter les fortes capacités de généralisation des modèles unimodaux pré-entraînés pour des tâches multimodales telles que le sous-titrage d'images. Deuxièmement, nous présentons une étude systématique de la robustesse des mesures d’évaluation des sous-titres d’images récemment proposées. Même si ces métriques correspondent bien aux jugements humains, nous avons constaté qu'elles ne sont pas robustes pour identifier les erreurs fines dans les légendes générées par le modèle. Il faut donc faire preuve de prudence lors de l'utilisation de ces métriques pour l'évaluation des sous-titres d'images. Nous espérons que nos résultats guideront de nouvelles améliorations dans l’évaluation automatique du sous-titrage d’images. / Image captioning is the artificial intelligence (AI) task of describing images in natural language. This AI task has several useful societal applications, such as accessibility for the visually impaired, automated content generation, human-robot interaction, and medical imaging analysis. Over the last eight years, image captioning research has seen tremendous progress in building strong models, collecting large scale datasets as well as developing automatic evaluation metrics. Despite such remarkable progress, image captioning research faces two major challenges: 1) How to build parameter-efficient models, and 2) How to build robust automatic evaluation metrics. In this thesis, we make contributions towards tackling each of these challenges. First, we propose a parameter efficient method (MAPL \cite{mapl}) that adapts pre-trained unimodal vision-only and language-only models for the multimodal task of image captioning. MAPL learns a lightweight mapping between the representation spaces of the unimodal models. Thus, MAPL can leverage the strong generalization capabilities of the pre-trained unimodal models for multimodal tasks such as image captioning. Second, we present a systematic study of the robustness of recently proposed image captioning evaluation metrics. Even though these metrics correlate well with human judgments, we found that these metrics are not robust in identifying fine-grained errors in model generated captions, and thus, caution needs to be exercised when using these metrics for image captioning evaluation. We hope our findings will guide further improvements in the automatic evaluation of image captioning.

Ce travail vise à contribuer à l’amélioration de notre connaissance de l’électrodynamique quantique des états liés, par des mesures de haute précision dans des états exotiques de la matière. Bien que notre connaissance de l’électrodynamique quantique des états liés et du problème relativiste à plusieurs corps aient fait des progrès importants ces dernières années, il reste des questions fondamentales dont la résolution nécessite d’augmenter le nombre et la précision des tests expérimentaux. La première partie de ce travail a été réalisée dans le cadre de l’expérience récente sur les ions muoniques d’hélium (μ4He+ et μ3He+) conduite par la collaboration CREMA. Cette expérience vise à fournir de nouvelles valeurs précises pour les rayons de charge moyens des noyaux des isotopes stables de l’hélium. Ces valeurs sont extraites de la mesure du déplacement de Lamb, c’est-à-dire, de la mesure de la différence d’énergie entre les États 2S − 2P. Une mesure des énergies de transition d’une précision d’au moins 50 ppm, permet de déterminer les rayons de charge des noyaux des isotopes d’hélium avec une incertitude de 0,03%, dix fois plus précise que les résultats précédents obtenus à partir de la diffusion d’électrons. La deuxième partie de ce travail a été de réaliser des mesures de haute précision de transitions de rayons X dans des ions fortement chargés, à l’aide d’un spectromètre à double cristal plan. Ces ions étaient produits dans le plasma d’une source d’ions de type ECRIS (electron-cyclotron resonance ion source). Le spectromètre utilisé est non seulement capable de fournir des mesures de haute précision, mais aussi des mesures sans référence à des énergies de transition théorique ou expérimentales. Quatre énergies de transition de n = 2 → n = 1 ont été mesurées pour des ions d’argon de trois états de charge différents, héliumoïde, lithiumoïde et berylliumoïde, avec une précision meilleure que 3 ppm. La largeur naturelle de chaque raie a également été obtenue expérimentalement. Les résultats trouvés sont en excellent accord avec les calculs théoriques les plus récents. / This work aims to provide insight on Bound-State Quantum-Electrodynamics (BSQED) by experimental fundamentals high-precision tests in exotic states of matter. Although BSQED and the relativistic many-body problem have been undergoing important progress, there are still some issues that require the increase of the number and accuracy of experimental fundamental tests. The first part of this work was done within the framework of the recent experiment in muonic helium ions (μ4He+ and μ3He+) by the CREMA collaboration. This experiment, aims to provide new accurate values for the root-mean-square (rms) charge radii of the helium isotopes nuclei that are extracted from the measurement of the Lamb Shift, i.e., the measurement of the energy difference between the 2S − 2P states. With the goal of measuring the transition energies with an accuracy of at least 50 ppm, the rms charge radii of the helium isotopes will be determined with an uncertainty of 0.03%, a factor of ten more precise than previous results obtained from electron scattering. The second part of this work aims the high-precision measurement of x-ray transitions in Highly-Charged Ions (HCI) using a Double-Crystal Spectrometer (DCS). These ions were produced in the plasma of an Electron-Cyclotron Resonance Ion Source (ECRIS). This kind of spectrometer is able not only to provide high-precision measurements but also reference-free measurements, without reference to any theoretical or experimental energy. Four transitions energies from n = 2 → n = 1 have been measured in an argon plasma in three different charge states, He-, Be- and Li-like, with an accuracy of better than 3 ppm. Besides the energies, the natural width of each transition has also been experimentally obtained. The obtained results are in excellent agreement with the most recent theoretical calculations.