Teses / dissertações sobre o tema "Multi-labels"

In the late 1960s a group of women became interested in forming a chapter of the National Organization for Women (NOW) in Richmond. These women, led by Zelda Nordlinger and Holt Carlton, followed a pragmatic, big-tent approach to women’s activism. This ideological and tactical openness defies traditional historical labels as these women fluidly moved through organizations and tactics in order to gain a stronger local following. Richmond’s NOW chapter, while staying attuned to the national organization’s platform, remained relatively autonomous and parochial in its tactics and pursuits. Further, Richmond NOW showed a marked change around 1974 with an influx of newer women into the organization. The Equal Rights Amendment (ERA) struggle provided the local movement with new prominence. With the interjection of new blood the chapter saw a shift in its tactics and policy. The newer cohort of women maintained a belief in a pragmatic, big-tent approach; however, they interpreted it differently. The chapter became more procedural and organizationally based. It also narrowed its focus and tactics, seeing the first generation’s free-wheeling style as a hindrance to organizational success. The different political experiences of these two cohorts led to different visions of Richmond’s NOW chapter.

The identification of the biomolecular interaction partners of small bioactive molecules is a fundamental problem in drug discovery and cell biology. This thesis describes the development of fluorescent chemical probes to identify the biomolecular targets of the known organophosphate toxin, phenyl saligenin phosphate (PSP), and the cardioprotective agent diazoxide. PSP is an organophosphate toxin that irreversibly inhibits hydrolase enzymes such as trypsin and chymotrypsin along with the common organophosphate target acetylcholine esterase. PSP is also suspected of affecting many other cell functions and may interact with a large number of cellular proteins. In this work phenyl saligenin phosphate has been synthesised and its inhibitory effect on the action of transglutaminase 2 (TGase2) demonstrated. Analogues of PSP containing an attached dansyl amide fluorescent group have been prepared and incubated with purified enzymes trypsin, chymotrypsin and TGase2. SDS-PAGE analysis demonstrates effective fluorescent labelling and a covalent interaction between the toxin analogue and the enzymes. The KATP channel opener, diazoxide displays marked cardioprotective effects and is reported to bind to mitochondrial KATP channels. However, the molecular structure of these channels is still largely unknown. This thesis describes the design and the synthesis of a chemical tool to covalently attach fluorescently labels to the proteins which will bind diazoxide. Chemical tools for fluorescent labelling of diazoxide targeted proteins have been prepared. Each consists of a photochemically activated reactive ‘barb’ and coupled fluorescent component linked to modified diazoxide bait. In developing these molecules, a range of functionalised diazoxide bait components were prepared and tested for retained biological activity compared to the parent compound. Two active analogues were linked to either benzophenone or diazirine (photoreactive) and dansyl amide (fluorescent) components. The non-specific photochemical reactivity of these labelling compounds with bovine serine albumin was established. The incubation and photolysis with mitochondrial extracts showed selective photo labelling of only three biomolecular components. The identification of these biomolecules is the subject of on-going investigation.

L'annotation des données est cruciale pour l'apprentissage automatique, notamment dans les domaines techniques, où la qualité et la quantité des données annotées affectent significativement l'efficacité des modèles entraînés. L'utilisation de personnel humain est coûteuse, surtout lors de l'annotation pour la classification multi-labels, les instances pouvant être associées à plusieurs labels. L'apprentissage actif (AA) vise à réduire les coûts d'annotation en sélectionnant intelligemment des instances pour l'annotation, plutôt que de les annoter de manière aléatoire. L'attention récente portée aux transformers a mis en lumière le potentiel de l'AA dans ce contexte. De plus, le mécanisme de fine-tuning, où seules quelques données annotées sont utilisées pour entraîner le modèle sur une nouvelle tâche, est parfaitement en accord avec l'objectif de l'AA de sélection des meilleures données à annoter. Nous étudions donc l'utilisation de l'AA dans le contexte des transformers pour la tâche de classification multi-labels. Hors, la plupart des stratégies AA, lorsqu'elles sont appliquées à ces modèles, conduisent à des temps de calcul excessifs, ce qui empêche leurs utilisations au cours d'une interaction homme-machine en temps réel. Afin de pallier ce problème, nous utilisons des stratégies d'AA plus rapides, basées sur l'incertitude. D'abord, nous mettons l'accent sur l'application de six stratégies d'AA différentes sur deux modèles transformers. Nos travaux mettent en évidence qu'un certain nombre de stratégies basées sur l'incertitude ne surpassent pas l'échantillonnage aléatoire lorsqu'elles sont appliquées aux modèles transformers. Afin d'évaluer si ces résultats sont dûs à un biais des stratégies basées sur l'incertitude, une approche de pré-clustering est introduite pour ajouter de la diversité dans la sélection des instances. Enfin, nous nous penchons sur les défis pratiques de la mise en œuvre de l'AA dans des contextes industriels. Notamment, l'écart entre les cycles de l'AA laisse du temps inutilisé aux annotateurs. Pour résoudre ce problème, nous étudions des méthodes alternatives de sélection d'instances, visant à maximiser l'efficacité de l'annotation en s'intégrant de manière transparente au processus de l'AA. Nous commençons par adapter deux méthodes existantes aux transformers, en utilisant respectivement un échantillonnage aléatoire et des informations de cycle d'AA périmées. Ensuite, nous proposons notre méthode novatrice basée sur l'annotation des instances pour rééquilibrer la distribution des labels. Notre approche atténue les biais, améliore les performances du modèle (jusqu'à 23% d'amélioration sur le score F1), limite les disparités dépendantes de la stratégie (diminution de près de 50% de l'écart-type) et réduit le déséquilibre des libellés (diminution de 30% du ratio moyen de déséquilibre). Nos travaux ravivent ainsi la promesse de l'AA en montrant que son intégration adaptée dans un projet d'annotation se traduit par une amélioration des performances du modèle final entraîné
Data annotation is crucial for machine learning, especially in technical domains, where the quality and quantity of annotated data significantly impact the effectiveness of trained models. Human annotation is costly, particularly for multi-label classification tasks, as instances may be associated with multiple labels. Active Learning (AL) aims to reduce annotation costs by intelligently selecting instances for annotation, rather than annotating randomly. Recent attention on transformers has highlighted the potential of AL in this context. Moreover, the fine-tuning mechanism, where only a few annotated data points are used to train the model for a new task, aligns well with the goal of AL to select the best data for annotation. We investigate the use of AL in the context of transformers for multi-label classification tasks. However, most AL strategies, when applied to these models, lead to excessive computational time, hindering their use in real-time human-machine interaction. To address this issue, we employ faster AL strategies based on uncertainty. First, we focus on applying six different AL strategies to two transformer models. Our work highlights that several uncertainty-based strategies do not outperform random sampling when applied to transformer models. To evaluate if these results stem from a bias in uncertainty-based strategies, we introduce a pre-clustering approach to add diversity to instance selection. Lastly, we tackle the practical challenges of implementing AL in industrial contexts. Particularly, the gap between AL cycles leaves idle time for annotators. To resolve this, we explore alternative instance selection methods aiming to maximize annotation efficiency by seamlessly integrating with the AL process. We start by adapting two existing methods to transformers, using random sampling and outdated AL cycle information, respectively. Then, we propose our innovative method based on instance annotation to rebalance label distribution. Our approach mitigates biases, improves model performance (up to 23% improvement on the F1 score), reduces strategy-dependent disparities (nearly 50% decrease in standard deviation), and decreases label imbalance (30% decrease in the mean imbalance ratio). Our work thus revives the promise of AL by demonstrating that its adapted integration into an annotation project results in improved performance of the final trained model

This thesis presents a real-time multi-face tracking system, which is able to track multiple faces for live videos, broadcast, real-time conference recording, etc. The real-time output is one of the most significant advantages. Our proposed tracking system is comprised of three parts: face detection, feature extraction and tracking. We deploy a three-layer Convolutional Neural Network (CNN) to detect a face, a one-layer CNN to extract the features of a detected face and a shallow network for face tracking based on the extracted feature maps of the face. The performance of our multi-face tracking system enables the tracker to run in real-time without any on-line training. This algorithm does not need to change any parameters according to different input video conditions, and the runtime cost will not be affected significantly by an the increase in the number of faces being tracked. In addition, our proposed tracker can overcome most of the generally difficult tracking conditions which include video containing a camera cut, face occlusion, false positive face detection, false negative face detection, e.g. due to faces at the image boundary or faces shown in profile. We use two commonly used metrics to evaluate the performance of our multi-face tracking system demonstrating that our system achieves accurate results. Our multi-face tracker achieves an average runtime cost around 0.035s with GPU acceleration and this runtime cost is close to stable even if the number of tracked faces increases. All the evaluation results and comparisons are tested with four commonly used video data sets.

Avec l’avènement et le développement rapide des technologies numériques, les données sont devenues à la fois un bien précieux et très abondant. Cependant, avec une telle profusion, se posent des questions relatives à la qualité et l’étiquetage de ces données. En effet, à cause de l’augmentation des volumes de données disponibles, alors que le coût de l’étiquetage par des experts humains reste très important, il est de plus en plus nécessaire de pouvoir renforcer l’apprentissage semi-supervisé grâce l’exploitation des données nonlabellisées. Ce problème est d’autant plus marqué dans le cas de l’apprentissage multilabels, et en particulier pour la régression, où chaque unité statistique est guidée par plusieurs cibles différentes, qui prennent la forme de scores numériques. C’est dans ce cadre fondamental, que s’inscrit cette thèse. Tout d’abord, nous commençons par proposer une méthode d’apprentissage pour la régression semi-supervisée, que nous mettons à l’épreuve à travers une étude expérimentale détaillée. Grâce à cette nouvelle méthode, nous présentons une deuxième contribution, plus adaptée au contexte multi-labels. Nous montrons également son efficacité par une étude comparative, sur des jeux de données issues de la littérature. Par ailleurs, la dimensionnalité du problème demeure toujours la difficulté de l’apprentissage automatique, et sa réduction suscite l’intérêt de plusieurs chercheurs dans la communauté. Une des tâches majeures répondant à cette problématique est la sélection de variables, que nous proposons d’étudier ici dans un cadre complexe : semi-supervisé, multi-labels et pour la régression
With the advent and rapid growth of digital technologies, data has become a precious asset as well as plentiful. However, with such an abundance come issues about data quality and labelling. Because of growing numbers of available data volumes, while human expert labelling is still important, it is more and more necessary to reinforce semi-supervised learning with the exploitation of unlabeled data. This problem is all the more noticeable in the multi-label learning framework, and in particular for regression, where each statistical unit is guided by many different targets, taking the form of numerical scores. This thesis focuses on this fundamental framework. First, we begin by proposing a method for semi-supervised regression, that we challenge through a detailed experimental study. Thanks to this new method, we present a second contribution, more fitted to the multi-label framework. We also show its efficiency with a comparative study on literature data sets. Furthermore, the problem dimension is always a pain point of machine learning, and reducing it sparks the interest of many researchers. Feature selection is one of the major tasks addressing this problem, and we propose to study it here in a complex framework : for semi-supervised, multi-label regression. Finally, an experimental validation is proposed on a real problem about automatic annotation of tires, to tackle the needs expressed by the industrial partner of this thesis

En classification multi-labels graduée (CMLG), chaque instance est associée à un ensemble de labels avec des degrés d’association gradués. Par exemple, une même molécule odorante peut être associée à une odeur forte ‘musquée’, une odeur modérée ‘animale’, et une odeur faible ‘herbacée’. L’objectif est d’apprendre un modèle permettant de prédire l’ensemble gradué de labels associé à une instance à partir de ses variables descriptives. Par exemple, prédire l’ensemble gradué d’odeurs à partir de la masse moléculaire, du nombre de liaisons doubles, et de la structure de la molécule. Un autre domaine intéressant de la CMLG est les systèmes de recommandation. En effet, les appréciations des utilisateurs par rapport à des items (produits, services, livres, films, etc) sont d’abord collectées sous forme de données MLG (l’échelle d’une à cinq étoiles est souvent utilisée). Ces données sont ensuite exploitées pour recommander à chaque utilisateur des items qui ont le plus de chance de l’intéresser. Dans cette thèse, une étude théorique approfondie de la CMLG permet de ressortir les limites des approches existantes, et d’assoir un ensemble de nouvelles approches apportant des améliorations évaluées expérimentalement sur des données réelles. Le cœur des nouvelles approches proposées est l’exploitation des relations entre les labels. Par exemple, une molécule ayant une forte odeur ‘musquée’ émet souvent une odeur faible ou modérée ‘animale’. Cette thèse propose également de nouvelles approches adaptées au cas des molécules odorantes et au cas des gros volumes de données collectées dans le cadre des systèmes de recommandation
In graded multi-label classification (GMLC), each instance is associated to a set of labels with graded membership degrees. For example, the same odorous molecule may be associated to a strong 'musky' odor, a moderate 'animal' odor, and a weak 'grassy' odor. The goal is to learn a model to predict the graded set of labels associated to an instance from its descriptive variables. For example, predict the graduated set of odors from the molecular weight, the number of double bonds, and the structure of the molecule. Another interesting area of the GMLC is recommendation systems. In fact, users' assessments of items (products, services, books, films, etc.) are first collected in the form of GML data (using the one-to-five star rating). These data are then used to recommend to each user items that are most likely to interest him. In this thesis, an in-depth theoretical study of the GMLC allows to highlight the limits of existing approaches, and to introduce a set of new approaches bringing improvements evaluated experimentally on real data. The main point of the new proposed approaches is the exploitation of relations between labels. For example, a molecule with a strong 'musky' odor often has a weak or moderate 'animal' odor. This thesis also proposes new approaches adapted to the case of odorous molecules and to the case of large volumes of data collected in the context of recommendation systems

En classification multi-labels graduée (CMLG), chaque instance est associée à un ensemble de labels avec des degrés d’association gradués. Par exemple, une même molécule odorante peut être associée à une odeur forte ‘musquée’, une odeur modérée ‘animale’, et une odeur faible ‘herbacée’. L’objectif est d’apprendre un modèle permettant de prédire l’ensemble gradué de labels associé à une instance à partir de ses variables descriptives. Par exemple, prédire l’ensemble gradué d’odeurs à partir de la masse moléculaire, du nombre de liaisons doubles, et de la structure de la molécule. Un autre domaine intéressant de la CMLG est les systèmes de recommandation. En effet, les appréciations des utilisateurs par rapport à des items (produits, services, livres, films, etc) sont d’abord collectées sous forme de données MLG (l’échelle d’une à cinq étoiles est souvent utilisée). Ces données sont ensuite exploitées pour recommander à chaque utilisateur des items qui ont le plus de chance de l’intéresser. Dans cette thèse, une étude théorique approfondie de la CMLG permet de ressortir les limites des approches existantes, et d’assoir un ensemble de nouvelles approches apportant des améliorations évaluées expérimentalement sur des données réelles. Le cœur des nouvelles approches proposées est l’exploitation des relations entre les labels. Par exemple, une molécule ayant une forte odeur ‘musquée’ émet souvent une odeur faible ou modérée ‘animale’. Cette thèse propose également de nouvelles approches adaptées au cas des molécules odorantes et au cas des gros volumes de données collectées dans le cadre des systèmes de recommandation
In graded multi-label classification (GMLC), each instance is associated to a set of labels with graded membership degrees. For example, the same odorous molecule may be associated to a strong 'musky' odor, a moderate 'animal' odor, and a weak 'grassy' odor. The goal is to learn a model to predict the graded set of labels associated to an instance from its descriptive variables. For example, predict the graduated set of odors from the molecular weight, the number of double bonds, and the structure of the molecule. Another interesting area of the GMLC is recommendation systems. In fact, users' assessments of items (products, services, books, films, etc.) are first collected in the form of GML data (using the one-to-five star rating). These data are then used to recommend to each user items that are most likely to interest him. In this thesis, an in-depth theoretical study of the GMLC allows to highlight the limits of existing approaches, and to introduce a set of new approaches bringing improvements evaluated experimentally on real data. The main point of the new proposed approaches is the exploitation of relations between labels. For example, a molecule with a strong 'musky' odor often has a weak or moderate 'animal' odor. This thesis also proposes new approaches adapted to the case of odorous molecules and to the case of large volumes of data collected in the context of recommendation systems

La représentation corporelle est l’ensemble des fonctions cognitives permettant le suivi de l’état du corps. Elle est impliquée dans des situations diverses, comme la perception des dimensions physiques du corps, l’action, ou encore la génération d’attitudes à propos du corps. Pour réaliser ces fonctions, elle se base de manière flexible sur un ensemble d’informations sensorimotrices, ainsi que sur les croyances, attentes et émotions de l’individu. Parmi les sources d’informations disponibles à propos du corps, l’influence sociale peut être un facteur de risque, de maintien, et de sévérité des distorsions de l’image du corps. Pourtant, si l’influence sociale sur les aspects attitudinaux de la représentation corporelle est bien établie, il existe peu d’évidence expérimentale de telles influences sur ses aspects perceptifs. Cette thèse a ainsi pour objectif d’étudier l’intégration d’informations sociales à la dimension perceptive de la représentation de la corpulence. Pour cela, nous avons réalisé une série d’expériences auprès de jeunes femmes, une catégorie de la population particulièrement sujette aux distorsions de la représentation corporelle. Un premier axe se focalise sur l'influence interpersonnelle en testant l'effet de labels de poids sur des jugements perceptifs. Pour étudier leur influence informationnelle, nous avons fait varier la fiabilité de plusieurs signaux pour étudier la manière dont elles étaient combinées. Nos résultats indiquent que les labels de poids ont une influence réduite sur les jugements de corpulence. Un second axe porte sur l'influence médiatique. La surexposition visuelle à certains types de corps est associée à l’insatisfaction corporelle, et pourrait contribuer à expliquer certaines distorsions perceptives et attitudinales de la représentation corporelle. Dans ce contexte, l’adaptation visuelle à des corps pourrait expliquer comment l’exposition prolongée à des corps minces peut mener à une surestimation de la corpulence propre. Nous avons testé certaines hypothèses de cette théorie adaptative des distorsions de l'image du corps. Ces expériences soulignent certaines limites de la théorie adaptative ; en particulier, il est incertain que les effets d’adaptation puissent influencer la représentation que les individus ont de leurs propres corps. En conclusion, nos résultats indiquent que la dimension perceptive de la représentation de la corpulence pourrait résister à certains types d’influences sociales interpersonnelles et médiatiques
Body representation is the set of cognitive functions that track the state of the body. It is involved in a variety of situations, such as the perception of the physical dimensions of the body, action, and the generation of attitudes towards the body. To perform these functions, it relies on the flexible use of a range of sensorimotor information, as well as on the individual's beliefs, expectations and emotions. Among the sources of information available about the body, social influence can be a risk, maintenance, and severity factor in body image distortions. However, while social influence on the attitudinal aspects of body representation is well established, there is little experimental evidence of such influence on its perceptual aspects. The aim of this thesis is to study the integration of social information into the perceptual dimension of the representation of body size. To this end, we conducted a series of experiments with young women, a demographic that is particularly prone to distortions of body representation. A first axis focuses on interpersonal influence by testing the effect of weight labels on perceptual judgments. To investigate their informational influence, we manipulated the reliability of multiple cues to study how they were combined. Our results indicate that weight labels have a limited influence on judgments of body size. A second axis focuses on another type of social influence, media influence. Visual overexposure to specific body types is associated with body dissatisfaction, and could help explain the perceptual and attitudinal distortions of body representation. In this context, visual adaptation to bodies could explain how prolonged exposure to thin bodies can lead to an overestimation of one's own body size. We tested some of the hypotheses of this adaptation theory of body image distortion. These experiments highlight some limitations of the adaptation account; in particular, it is uncertain whether adaptation effects can influence the representation that individuals have of their own bodies. In conclusion, our results suggest that the perceptual dimension of the representation of body size may be resistant to some types of interpersonal and media social influence

Nowadays, imaging of the human brain is vastly used in clinical settings and by the neuroscientific research community. There is an ever-increasing demand for novel biomedical image analysis approaches and tools to study the brain from its early intrauterine stage through adolescence to adulthood. The intrauterine period, in particular, is a crucial stage for the study of early neurodevelopmental processes. The idiosyncratic nature of the fetal brain poses numerous challenges and asks for the development of new techniques that take into consideration the peculiarities of in utero neurodevelopment. Although still in its infancy, medical image analysis techniques are progressively landing on the study of fetal brains. The purpose of this thesis is to develop automatic segmentation approaches that can be applied to brains at different life stages, including the gestational period, and investigate in utero brain development under ventriculomegaly.
En la actualidad, las imagenes del cerebro humano son ampliamente utilizadas en entornos clıınicos y por la comunidad neurocientııfica. Existe una demanda, cada vez mayor, de herramientas y enfoques de analisis de imagenes biomédicas novedosos para estudiar el cerebro desde su temprana etapa intrauterina hasta la adolescencia y la edad adulta. El periodo intrauterino, en particular, es una etapa crucial para el estudio de los procesos iniciales del neurodesarrollo. La naturaleza idiosincrasica del cerebro fetal plantea numerosos desafııos y requiere el desarrollo de nuevas técnicas que contemplen las peculiaridades del neurodesarrollo fetal. Aunque todavııa esta en su infancia, las técnicas de analisis de imagenes médicas estan llegando progresivamente al estudio de los cerebros fetales. El objetivo de esta tesis es desarrollar métodos automaticos de segmentación que puedan aplicarse a cerebros en distintas etapas de la vida, incluyendo el periodo gestacional, e investigar el desarrollo del cerebro fetal con ventriculomegalia.

Text Classification (TC) which is an important field in information technology has many valuable applications. When facing the sea of information resources, the objects of TC are more complicated and diversity. The researches in pursuit of effective and practical TC technology are fairly challenging. More and more researchers regard that multi-label TC is more suited for many applications. This thesis analyses the difficulties and problems in multi-label TC and Chinese text representation based on a mass of algorithms for single-label TC and multi-label TC. Aiming at high dimensionality in feature space, sparse distribution in text representation and poor performance of multi-label classifier, this thesis will bring forward corresponding algorithms from different angles.Focusing on the problem of dimensionality “disaster” when Chinese texts are represented by using n-grams, two-step feature selection algorithm is constructed. The method combines filtering rare features within class and selecting discriminative features across classes. Moreover, the proper value of “n”, the strategy of feature weight and the correlation among features are discussed based on variety of experiments. Some useful conclusions are contributed to the research of n-gram representation in Chinese texts.In a view of the disadvantage in Latent Dirichlet Allocation (LDA) model, that is, arbitrarily revising the variable in smooth process, a new strategy for smoothing based on Tolerance Rough Set (TRS) is put forward. It constructs tolerant class in global vocabulary database firstly and then assigns value for out-of-vocabulary (oov) word in each class according to tolerant class.In order to improve performance of multi-label classifier and degrade computing complexity, a new TC method based on LDA model is applied for Chinese text representation. It extracts topics statistically from texts and then texts are represented by using the topic vector. It shows competitive performance both in English and in Chinese corpus.To enhance the performance of classifiers in multi-label TC, a compound classification framework is raised. It partitions the text space by computing the upper approximation and lower approximation. This algorithm decomposes a multi-label TC problem into several single-label TCs and several multi-label TCs which have less labels than original problem. That is, an unknown text should be classified by single-label classifier when it is partitioned into lower approximation space of some class. Otherwise, it should be classified by corresponding multi-label classifier.An application system TJ-MLWC (Tongji Multi-label Web Classifier) was designed. It could call the result from Search Engines directly and classify these results real-time using improved Naïve Bayes classifier. This makes the browse process more conveniently for users. Users could locate the texts interested immediately according to the class information given by TJ-MLWC
La thèse est centrée sur la Classification de texte, domaine en pleine expansion, avec de nombreuses applications actuelles et potentielles. Les apports principaux de la thèse portent sur deux points : Les spécificités du codage et du traitement automatique de la langue chinoise : mots pouvant être composés de un, deux ou trois caractères ; absence de séparation typographique entre les mots ; grand nombre d’ordres possibles entre les mots d’une phrase ; tout ceci aboutissant à des problèmes difficiles d’ambiguïté. La solution du codage en «n-grams »(suite de n=1, ou 2 ou 3 caractères) est particulièrement adaptée à la langue chinoise, car elle est rapide et ne nécessite pas les étapes préalables de reconnaissance des mots à l’aide d’un dictionnaire, ni leur séparation. La classification multi-labels, c'est-à-dire quand chaque individus peut être affecté à une ou plusieurs classes. Dans le cas des textes, on cherche des classes qui correspondent à des thèmes (topics) ; un même texte pouvant être rattaché à un ou plusieurs thème. Cette approche multilabel est plus générale : un même patient peut être atteint de plusieurs pathologies ; une même entreprise peut être active dans plusieurs secteurs industriels ou de services. La thèse analyse ces problèmes et tente de leur apporter des solutions, d’abord pour les classifieurs unilabels, puis multi-labels. Parmi les difficultés, la définition des variables caractérisant les textes, leur grand nombre, le traitement des tableaux creux (beaucoup de zéros dans la matrice croisant les textes et les descripteurs), et les performances relativement mauvaises des classifieurs multi-classes habituels
文本分类是信息科学中一个重要而且富有实际应用价值的研究领域。随着文本分类处理内容日趋复杂化和多元化，分类目标也逐渐多样化，研究有效的、切合实际应用需求的文本分类技术成为一个很有挑战性的任务，对多标签分类的研究应运而生。本文在对大量的单标签和多标签文本分类算法进行分析和研究的基础上，针对文本表示中特征高维问题、数据稀疏问题和多标签分类中分类复杂度高而精度低的问题，从不同的角度尝试运用粗糙集理论加以解决，提出了相应的算法，主要包括：针对n-gram作为中文文本特征时带来的维数灾难问题，提出了两步特征选择的方法，即去除类内稀有特征和类间特征选择相结合的方法，并就n-gram作为特征时的n值选取、特征权重的选择和特征相关性等问题在大规模中文语料库上进行了大量的实验，得出一些有用的结论。针对文本分类中运用高维特征表示文本带来的分类效率低，开销大等问题，提出了基于LDA模型的多标签文本分类算法，利用LDA模型提取的主题作为文本特征，构建高效的分类器。在PT3多标签分类转换方法下，该分类算法在中英文数据集上都表现出很好的效果，与目前公认最好的多标签分类方法效果相当。针对LDA模型现有平滑策略的随意性和武断性的缺点，提出了基于容差粗糙集的LDA语言模型平滑策略。该平滑策略首先在全局词表上构造词的容差类，再根据容差类中词的频率为每类文档的未登录词赋予平滑值。在中英文、平衡和不平衡语料库上的大量实验都表明该平滑方法显著提高了LDA模型的分类性能，在不平衡语料库上的提高尤其明显。针对多标签分类中分类复杂度高而精度低的问题，提出了一种基于可变精度粗糙集的复合多标签文本分类框架，该框架通过可变精度粗糙集方法划分文本特征空间，进而将多标签分类问题分解为若干个两类单标签分类问题和若干个标签数减少了的多标签分类问题。即，当一篇未知文本被划分到某一类文本的下近似区域时，可以直接用简单的单标签文本分类器判断其类别；当未知文本被划分在边界域时，则采用相应区域的多标签分类器进行分类。实验表明，这种分类框架下，分类的精确度和算法效率都有较大的提高。本文还设计和实现了一个基于多标签分类的网页搜索结果可视化系统（MLWC），该系统能够直接调用搜索引擎返回的搜索结果，并采用改进的Naïve Bayes多标签分类算法实现实时的搜索结果分类，使用户可以快速地定位搜索结果中感兴趣的文本。

Les réseaux neuronaux ont fait preuve d'avancées remarquables dans des tâches réputées difficiles, comme la classification automatique d'images ou le traitement du langage naturel. Toutefois, leur nature de boîte noire limite leur explicabilité et entrave leur capacité à exploiter des connaissances extérieures aux données. Cette thèse a pour but d'explorer et de proposer des techniques d'intégration des connaissances de haut niveau dans les réseaux neuronaux afin d'améliorer les performances et l'interprétabilité. La première partie de la thèse est centrée sur l'intégration de connaissances aux données d'entrée d'un réseau. Son premier chapitre s'adresse à la préparation des données. On y propose une formalisation du prétraitement afin de garantir la transparence et la reproductibilité de cette étape. Cette formalisation nous permet d'étudier l'impact de la data-augmentation : pour caractériser ce qu'est une bonne préparation des données, et l'état informatif d'un dataset, un ensemble de mesures et de principes est proposé, ensuite des protocoles expérimentaux sont conçus afin d'évaluer ces principes sur le dataset BreakHis. Le deuxième chapitre de cette partie s'adresse à l'exploitation de connaissances haut-niveau pour l'établissement d'un ordre de présentation des données au réseau. Nous introduisons l'apprentissage par curriculum incrémental sur l'ordre de passage des données en entrée. Les résultats obtenus améliorent l'exactitude et la vitesse de convergence. Bien que cette étude soit menée sur le dataset BreakHis, nous pensons qu'elle est généralisable à n'importe quel autre dataset. La deuxième partie est centrée sur l'intégration de connaissances au sein de l'architecture du réseau et au niveau de sa sortie. Dans ce cadre, nous nous sommes intéressés à la classification multi-label hiérarchique, pour laquelle nous avons formalisé les connaissances représentant le lien hiérarchique. Pour cela nous avons introduit deux contraintes : l'une représentant le fait qu'un objet ne peut être affecté qu'à une seule classe à un niveau donné de la hiérarchie, et l'autre imposant que l'affectation globale d'un objet respecte la hiérarchie de classe (par exemple, on interdit de classer un élément comme abeille pour son sous-type et mammifère pour son sur-type). Nous avons conçu une architecture et une fonction de perte qui imposent ces deux contraintes durant l'apprentissage. L'architecture se distingue de l'état de l'art par le fait qu'un seul réseau est utilisé pour prédire simultanément les labels des différents niveaux : toutes les couches sont responsables de la prédiction du n-uplet des classes. Plusieurs variantes du réseau ont été expérimentées sur cinq jeux de données différents et les résultats confirment l'efficacité des contraintes hiérarchiques soutenant ainsi l'importance de la prise en compte de connaissances externes. Afin de raffiner les résultats de cette classification hiérarchique, nous avons introduit un mécanisme d'abstention, sous forme d'une troisième contrainte poussant le réseau à donner une prédiction au niveau de spécificité le plus précis possible sur lequel sa confiance est suffisante et s'abstenir sinon. Nous avons défini différents seuils de confiance et proposé différentes contraintes sur les seuils relativement à la hiérarchie des classes. Pour évaluer ce mécanisme, de nouvelles métriques de classification prenant en compte l'abstention ont été définies. Nous avons mené des expérimentations sur les mêmes jeux de données et les résultats ont montré l'intérêt de l'abstention, et la nécessité de définir un seuil empirique adapté à chaque dataset. Pour conclure, les travaux de cette thèse soulignent l'intérêt d'exploiter des connaissances externes dans le domaine des réseaux de neurones ceci au niveau des trois composantes de ce système d'apprentissage : en entrée pendant la préparation des données, dans la structure du réseau, et à la sortie lors de la prise de décision de classification
Neural networks have made remarkable improvements in challenging tasks such as automatic image classification and natural language processing. However, their black-box nature hinders explainability and limits their ability to leverage external knowledge. The purpose of this thesis is to explore and propose techniques for integrating knowledge into neural networks in order to improve their performance and interpretability. The first part of the thesis focuses on integrating knowledge at the input level. The first chapter deals with data preparation. A formalization of pre-processing is proposed to ensure the transparency and reproducibility of this step. This formalization enables us to study the impact of data augmentation: to characterize a good data preparation, and the informative state of a dataset, a set of measures and principles is proposed, then experimental protocols are designed to evaluate these principles on the BreakHis dataset. The second chapter of this part focuses on exploiting high-level knowledge to determine the order in which data should be inserted into the network. We introduce an incremental curriculum learning for ordering the input data. The results obtained show an improvement of accuracy and convergence speed. Although this study is carried out on the BreakHis dataset, we believe that it can be generalized to any other dataset. The second part is devoted to the integration of knowledge within the network architecture and at the output level. In this context, we focus on hierarchical multi-label classification, for which we formalize the knowledge representing the hierarchical link. For this aim, we introduce two constraints: one representing the fact that an object can only be assigned to one class at a given level of the hierarchy, and the other imposing that the global assignment of an object respects the class hierarchy (for example, we forbid classifying an element as a bee for its sub-type and a mammal for its super-type). We design an architecture and a loss function that impose these two constraints during learning. The architecture differs from the state of the art in that a single network is used to simultaneously predict the labels of the different levels: all layers are responsible for predicting the tuple of classes. Several variants of the network have been tested on five different datasets and the results confirm the efficiency of the hierarchical constraints, thus supporting the importance of taking external knowledge into account. In order to refine the results of this hierarchical classification, we introduce an abstention mechanism, in the form of a third constraint that enforces the network to give a prediction at the most precise level of specificity on which its confidence is sufficient and to abstain otherwise. We define different confidence thresholds and proposed different constraints on the thresholds accordingly to the class hierarchy. To evaluate this mechanism, new classification metrics that take abstention into account are defined. We carry out experiments on the same five datasets and the results show the interest of abstention, and the need to define empirical thresholds adapted to each dataset. In conclusion, the work in this thesis highlights the value of exploiting external knowledge, this is true for the three main components of a neural network: at the input level during data preparation, in the structure of the network, and at the output level when classification decisions are made

Pulsed electron-electron double resonance (PELDOR) is an emerging technique for nanometre distance measurements in nano-sized assemblies and between specific sites of molecules. Most commonly nitroxide radicals are used as probes for EPR distance measurements because they are easy to introduce in biological systems such as soluble and membrane proteins or nucleic acids. PELDOR distance measurements currently rely on data processing software which has been proven to accurately extract inter-spin distances from the dipolar coupling between two paramagnetic centres. However, when the dipolar coupling is affected by contributions from other close-by unpaired electrons inaccuracies as broadening effects and artefacts are introduced in the distance distributions derived. This challenge, commonly referred as multi-spin effects, has been affecting the extraction of accurate distance information from PELDOR measurements in chemical and biological systems with multiple spin labels. The aim of this project is to approach, identify and suppress inaccuracies introduced in PELDOR-based distance distributions by multi-spin effects. This is achieved through the synthesis of multiply labelled model systems which would allow for assessment of the impact of multi-spin effects on distance measurements of simple geometries whose behaviour can be easily predicted and modelled. In this work existing methods for suppression of multi-spin effects are tested, together with their efficiency and limitations. The results are used to devise better sets of parameters including alternative settings for extraction of accurate distances from multi-spin systems and to explore their efficiency and limitations. Additional effects influencing distance measurements by pulsed EPR are also examined; in particular the effects of orientation selection and their interplay with multi-spin effects is studied in depth. Studies on rigid symmetric and asymmetric chemical model systems together with heptameric channel membrane proteins allow for outlining of recommendations for PELDOR distance measurements settings on systems presenting similar structural features, including symmetries and inter-spin distances.

碩士
元智大學
資訊工程學系
105
With the development of society, more and more attention has been paid to psychiatric health and psychiatric illness. When people are depressed or suffer from psychiatric illness, it is an efficient and effective way to seek help from the Internet to help them to alleviate their suffering. Many psychiatric health websites have established forums and blogs to help people share their psychological problems with other users and psychologists. Other users and psychologists can give advice on how to respond to these psychological problems. Psychiatric health Web site has accumulated a large number of descriptions of psychological illness, which contains a wealth of emotion labels to express different psychological illness. Automatic identification of these mental illness labels can make online mental health services more efficient. In this paper, we propose a combined depth neural network framework BLSTM_CNN model to extract features from the text automatically. BLSTM is used to extract the sentence for each word, CNN is used to extract local features in a word, through the combination of BLSTM and CNN, according to the different emotion labels, implied more useful features can be extracted. The experimental results show that the BLSTM_CNN model is better than the CNN, LSTM and LSTM_CNN models.

博士
國立臺灣科技大學
資訊工程系
107
The problem of semi-supervised classification with non-uniformly distributed incomplete labels is frequently encountered in real world applications. The lack of positive information, the absence of negative examples and the non-uniform distribution of missing labels lead to the diminished accuracy of multi-label classification results. In this research, we propose the Semi-supervised Incomplete Training Label Recovery (SITLR) algorithm to solve the semi-supervised multi-label classification with incompletely labeled training data. With the proposed weight adjustment step and negative information initialization with LF-CARS algorithm, SITLR focuses on enhancing the information of labeled training instances according to the distribution of data, where it only recovers some important labels and the recovered training data can be applied to any existing multi-label classification algorithm for building a better classification model and generating better label predictions in the testing phase. The experiments verified the effectiveness of SITLR.

博士
國立成功大學
電腦與通信工程研究所
105
Internet protocol (IP) is the most widely used protocol for high-bandwidth data transmission and it has been thought as a solution to provide different high-quality services in the future. As the internet traffic increases rapidly, the network size is extended. Multi-protocol label switching (MPLS) is proposed to reduce the IP processing time because only label is processed during the packet transmission between nodes. Although MPLS partially releases the burden of IP network, packet routing still faces a bottleneck when the number of users is large. Optical packet switching (OPS) overcomes this difficulty by simplifying several layers into IP over optical network. To implement MPLS over optical work, optical codes (OC) are used as labels for packet switching in Generalized MPLS (GMPLS) network. Among several label approaches, spectral amplitude coding (SAC) lowers system complexity and is compatible with label stacking. The label of an optical pocket is composed by different wavelength components, which are encoded according to a signature code pattern. However, due to the incoherent property of light source, the phase intensity induced noise (PIIN) appears at the forwarding node when the optical code label is de-coded. PIIN cannot simply removed by increasing the signal power because its value is proportional to the detected optical current. Therefore, we design three optical code labelling (OCL) scenarios, to increase the probability of correctly decoding the label in core nodes (CNs). Since the packet is sent to the appropriate path, the label error rate (LER) at edge node (EN) is decreased. In the first approach, stuffed quadratic congruence code (SQC code) is proposed for optical label implementing. Because of its low cross-correlation value, the effect of PIIN can be decreased significantly. If the label can be decoded correctly, the forward node would generate proper control signal to direct the packet to a suitable path. This reduces the probability of packet missing and lowers the value of LER when optical packet is de-modulated at the end node. For the case of label stacking, labels with SQC codes can provide greater system improvements. To meet practical applications, the relation between SAC-labels and optical MPLS network performance is also analyzed in this dissertation by numerical simulation. In the second approach, a hybrid label for optical packet switching in GMPLS network is proposed by combining SAC optical code-division multiple access (OCDMA) with wavelength division multiplexing (WDM). The author considers two label assignment scenarios. Hybrid labels are sequentially assigned to path segments in a label switching path (LSP) based on code index or wavelength index. LER performance of these two label assignment scenarios are also analyzed. Better LER results is achieved by sequential wavelength assignment, due to the similar label numbers among wavelengths. Furthermore, the optimal channel number is derived to minimize the LER under a specific number of stacked labels. In the final approach, bipolar OCL is employed in GMPLS network to improve the efficiency of label-recognition and network throughput. Label switching capabilities in LER is greatly reduced since the proposed bipolar OCL enlarges the Hamming distance of the star diagram of the decoded label signals. The proposed label mapping mechanism is also achieved through SAC in physical layer. In performance analysis, a numerical simulation of LER is presented to quantify the switching efficiency. Results show the proposed bipolar coding technique reduces LER in switching process, resulting in an extension of LSP in GMPLS core network.

The objective of cross-modal retrieval is to retrieve relevant items from one modality (say image), given a query from another modality (say textual document). Cross-modal retrieval has various applications like matching image-sketch, audio-visual, near infrared-RGB, etc. Different feature representations of the two modalities, absence of paired correspondences, etc. makes this a very challenging problem. In this thesis, we have extensively looked at the cross-modal retrieval problem from different aspects and proposed methodologies to address them. • In the first work, we propose a novel framework, which can work with unpaired data of the two modalities. The method has two-steps, consisting of a hash code learning stage followed by a hash function learning stage. The method can also generate unified hash representations in post-processing stage for even better performance. Finally, we investigate, formulate and address the cross-modal hashing problem in presence of missing similarity information between the data items. • In the second work, we investigate how to make the cross-modal hashing algorithms scalable so that it can handle large amounts of training data and propose two solutions. The first approach builds on mini-batch realization of the previously formulated objective and the second is based on matrix factorization. We also investigate whether it is possible to build a hashing based approach without the need to learn a hash function as is typically done in literature. Finally, we propose a strategy so that an already trained cross-modal approach can be adapted and updated to take into account the real life scenario of increasing label space, without retraining the entire model from scratch. • In the third work, we explore semi-supervised approaches for cross-modal retrieval. We first propose a novel framework, which can predict the labels of the unlabeled data using complementary information from the different modalities. The framework can be used as an add-on with any baseline cross-modal algorithm. The second approach estimates the labels of the unlabeled data using nearest neighbor strategy, and then train a network with skip connections to predict the true labels. • In the fourth work, we investigate the cross-modal problem in an incremental multiclass scenario, where new data may contain previously unseen categories. We propose a novel incremental cross-modal hashing algorithm, which can adapt itself to handle incoming data of new categories. At every stage, a small amount of old category data termed exemplars is used, so as not to forget the old data while trying to learn for the new incoming data. • Finally, we investigate the effect of label corruption on cross-modal algorithms. We first study the recently proposed training paradigms, which focuses on small loss samples to build noise-resistant image classification models and improve upon that model using techniques like self-supervision and relabeling of large loss samples. Next we extend this work for cross-modal retrieval under noisy data.