Depuis des années, on observe un déclin des insectes pollinisateurs. La perte d’habitats naturels, en partie liée àl’urbanisation, est considérée comme l’une des causes majeures de ce déclin. Des populations d’abeilles(Hymenoptera : Anthophila) se maintiennent cependant en milieux urbains. La structure de leurs communautés,ainsi que leurs comportements de butinage et de nidification peuvent être affectés par les perturbations liées àl’urbanisation. Notre objectif était d’évaluer l’ampleur de ces modifications et de comprendre leurs mécanismes enutilisant 24 sites dans le Grand Lyon localisés selon un gradient d’urbanisation croissante (mesurée par laproportion de surfaces imperméables). Nous avons analysé les réseaux d’interactions pour étudier les relationsflore-abeilles et nous avons mis en place des aménagements pour la nidification (carrés de sol et hôtels à abeilles)sur 16 sites urbains ou périurbains pour étudier la dynamique de nidification et son impact sur les populations.Nous avons capturé 16352 spécimens appartenant à 293 espèces, soit près du tiers de la faune françaised’abeilles. Les sites périurbains (avec environ 50% de surface imperméable) avaient la plus grande richessespécifique. Les abeilles à langue longue et les abeilles cavicoles étaient plus présentes dans les milieux urbanisés,mais la spécialisation des interactions plantes-abeilles est restée stable le long du gradient d’urbanisation. Ladiversité spécifique des abeilles était associée de façon positive avec la diversité florale, la durée de floraison desespèces végétales et/ou leur floribondité. Les communautés d’abeilles étaient dépendantes de ces trois facteurs etaussi des plantes spontanées (natives ou naturalisées) plus que des plantes horticoles (ornementales ou exotiques).Enfin, les aménagements pour la nidification ont été colonisés par une faune d’abeilles diversifiée. Hormis Osmiabicornis, les abeilles ne présentaient aucune préférence quant aux substrats dans lesquels elles nidifiaient. Même sil’urbanisation change la structure des communautés d’abeilles, nos résultats confirment qu’une diversité importanted’abeilles sauvages peut perdurer dans des milieux moyennement, mais aussi fortement urbanisés.Dans un contexte d’urbanisation croissante et de déclin des abeilles, il semble indispensable de mettre en placedes plans de gestion en faveur de ces insectes pollinisateurs sauvages en agissant sur l’étendue des surfaces fleuries,les continuités écologiques entre ces surfaces, et une gestion plus appropriée des parcs, jardins et espaces vertsurbains. La présence d’une grande diversité d’espèces y compris dans des milieux très anthropisés fait vraiment desabeilles un groupe phare pour sensibiliser les citoyens à la biodiversité et aux services écosystémiques / Evidence has been accumulating for years that pollinator populations are declining. The loss of natural habitats,in part linked to urbanization, is considered to be one of the major causes of this decline. Some bee populationspersist nevertheless in urban environments. The structure of their communities, as well as their foraging and nestingbehaviors can be affected by urbanization. Our objective was to assess the magnitude of these changes and tounderstand their mechanisms by using 24 sites located in the Grand Lyon along a gradient of urbanization(measured by the proportion of impervious surface). We analyzed interaction networks to investigate plant-beerelations, and man-made nesting structures on urban sites (squares of soil and bee hotels) to study nesting dynamicsand its relevance for bee populations.We collected 16,352 specimens belonging to 293 species. Periurban environments (i.e., with ca. 50% ofimpervious surface) had the greatest diversity of bees. Long-tongued bees and cavity nesting bees were moreprevalent in urban environments, but the specialization of plant-bee interactions remained stable along the gradientof urbanization. The species richness of the bee community was positively associated with floral diversity, theduration of flowering of plant species and/or their floribundity. Bee communities depended on these three factorsand also spontaneous (native and naturalized) plant species more than on horticultural plants (ornamentals andexotics). Finally, man-made nesting sites were colonized by a diverse bee fauna. Apart from Osmia bicornis, beesshowed no preference for the substrates in which they nested. Our work confirms that, even if urbanization changesthe structure of the bee communities, an important diversity of bee species can persist in periurban, but also urbanenvironments.In a context of increasing urbanization and declining bee populations, it appears essential to create managementplans for these wild pollinators by acting on the surface of flowering areas, the ecological network linking them,and the appropriate management of parks, gardens, and recreational areas. The presence of a diverse array of beespecies even in the most urbanized area makes these pollinators worthy of being a flagship group to raise theawareness of citizens about biodiversity and ecosystemic services

Submitted by Sandra Maria Neri Santiago (sandra.neri@ufpe.br) on 2016-04-18T18:41:44Z No. of bitstreams: 2 license_rdf: 1232 bytes, checksum: 66e71c371cc565284e70f40736c94386 (MD5) dissertação.pdf: 1661681 bytes, checksum: a525e39e828632534b1a6e345c1ba890 (MD5) / Made available in DSpace on 2016-04-18T18:41:44Z (GMT). No. of bitstreams: 2 license_rdf: 1232 bytes, checksum: 66e71c371cc565284e70f40736c94386 (MD5) dissertação.pdf: 1661681 bytes, checksum: a525e39e828632534b1a6e345c1ba890 (MD5) Previous issue date: 2005-08-29 / Este trabalho aborda o problema de transpor ferramentas e conceitos cartográficos para aplicações de realidade virtual, focando o uso da generalização cartográfica em mundos urbanos virtuais para turismo. A Generalização Cartográfica é uma área de Cartografia utilizada para se obter versões de mapas cartográficos. Em mundos virtuais este processo pode ser implementado com a geração de níveis de detalhamento (Level of Detail – LOD), mas não existem regras precisas para construir esses níveis de forma automática; grande parte dos trabalhos encontrados na literatura foca apenas na simplificação de polígonos. Para se obter versões generalizadas de mapas, a cartografia faz uso de doze operadores, aplicados pelo Cartógrafo segundo o seu conhecimento do problema. Estes operadores são: simplificar, suavizar, agregar, amalgamar, “merging”, colapsar, refinar, exagerar, aumentar, deslocar, classificação e simbolização. É realizada uma análise destes operadores verificando a viabilidade de sua aplicação em mundos virtuais. O conhecimento de um especialista na área de Cartografia é modelado através de regras de conhecimento, as quais alimentam um sistema especialista utilizado para selecionar os objetos de acordo com o tema, que nesta dissertação é o turismo, e para a aplicação dos operadores sobre o mundo virtual. Três dos operadores cartográficos foram escolhidos para serem implementados para mundos virtuais, pois já eram utilizados de forma isolada: simplificar, suavizar e simbolização. A partir destas informações, o SisGen3D, Sistema de Generalização para mundos virtuais construídos em VRML (Virtual Reality Modelling Language), é apresentado e validado com um mundo virtual, o Recife Antigo Virtual. Este sistema foi desenvolvido em JAVA, modelado com UML (Unified Modelling Language) e utiliza o gerenciador de banco de dados MySQL. / This work deals with the transposition of cartographic concepts and tools to virtual reality applications, using cartographic generalization in virtual urban worlds for tourism applications. Cartographic Generalization is an area in Cartography concerned with the generation of versions of cartographic maps. In virtual worlds, it can be implemented using Levels of Detail - LODs, but there are no known rules for constructing these levels automatically; most of the literature concentrates in polygon simplification. To get versions of maps, Cartography uses twelve operators, applied by the cartographer using domain-specific knowledge: simplification, smoothing, aggregation, amalgamation, merging, collapse, refinement, exaggeration, enhancement, displacement, classification and symbolization. An analysis of these operators is performed in the context of virtual reality. The cartographer domain-specific knowledge is modelled in knowledge rules in an expert system. This expert system is responsible for the selection of important objects and of the application of cartographic operators in virtual worlds. Three operators were implemented to virtual worlds: simplification, smoothing and symbolization. Finally, SisGen3D, a system for generalization virtual VRML worlds is presented. This system, developed in Java, modelled in UML and using the MySQL database, is validated using the Recife Antigo Virtual as a testbed.

O processo de mineração de dados tem como objetivo encontrar o conhecimento implícito em um conjunto de dados para auxiliar a tomada de decisão. Do ponto de vista do usuário, vários problemas podem ser encontrados durante a etapa de pós-processamento e disponibilização do conhecimento extraído, como a enorme quantidade de padrões gerados por alguns algoritmos de extração e a dificuldade na compreensão dos modelos extraídos dos dados. Além do problema da quantidade de regras, os algoritmos tradicionais de regras de associação podem levar à descoberta de conhecimento muito específico. Assim, pode ser realizada a generalização das regras de associação com o intuito de obter um conhecimento mais geral. Neste projeto é proposta uma metodologia interativa que auxilie na avaliação de regras de associação generalizadas, visando melhorar a compreensibilidade e facilitar a identificação de conhecimento interessante. Este auxílio é realizado por meio do uso de técnicas de visualização em conjunto com a aplicação medidas de avaliação objetivas e subjetivas, que estão implementadas no módulo de visualização de regras de associação generalizados denominado RulEE-GARVis, que está integrado ao ambiente de exploração de regras RulEE (Rule Exploration Environment). O ambiente RulEE está sendo desenvolvido no LABIC-ICMC-USP e auxilia a etapa de pós-processamento e disponibilização de conhecimento. Neste contexto, também foi objetivo deste projeto de pesquisa desenvolver o Módulo de Gerenciamento do ambiente de exploração de regras RulEE. Com a realização do estudo dirigido, foi possível verificar que a metodologia proposta realmente facilita a compreensão e a identificação de regras de associação generalizadas interessantes / The data mining process aims at finding implicit knowledge in a data set to aid in a decision-making process. From the users point of view, several problems can be found at the stage of post-processing and provision of the extracted knowledge, such as the huge number of patterns generated by some of the extraction algorithms and the difficulty in understanding the types of the extracted data. Besides the problem of the number of rules, the traditional algorithms of association rules may lead to the discovery of very specific knowledge. Thus, the generalization of association rules can be realized to obtain a more general knowledge. In this project an interactive methodology is proposed to aid in the evaluation of generalized association rules in order to improve the understanding and to facilitate the identification of interesting knowledge. This aid is accomplished through the use of visualization techniques along with the application of objective and subjective evaluation measures, which are implemented in the visualization module of generalized association rules called RulEE-GARVis, which is integrated with the Rule Exploration Environment RulEE. The RulEE environment is being developed at LABIC-ICMC-USP and aids in the post-processing and provision of knowledge. In this context, it was also the objective of this research project to develop the Module Management of the rule exploration environment RulEE. Through this directed study, it was verified that the proposed methodology really facilitates the understanding and identification of interesting generalized association rules

Les techniques d'apprentissage profond ont fait un bond au cours des dernières années, et ont considérablement changé la manière dont les tâches de traitement automatique du langage naturel (TALN) sont traitées. En quelques années, les réseaux de neurones et les plongements de mots sont rapidement devenus des composants centraux à adopter dans le domaine. La supervision distante (SD) est une technique connue en TALN qui consiste à générer automatiquement des données étiquetées à partir d'exemples partiellement annotés. Traditionnellement, ces données sont utilisées pour l'entraînement en l'absence d'annotations manuelles, ou comme données supplémentaires pour améliorer les performances de généralisation. Dans cette thèse, nous étudions comment la supervision distante peut être utilisée dans un cadre d'un TALN moderne basé sur l'apprentissage profond. Puisque les algorithmes d'apprentissage profond s'améliorent lorsqu'une quantité massive de données est fournie (en particulier pour l'apprentissage des représentations), nous revisitons la génération automatique des données avec la supervision distante à partir de Wikipédia. On applique des post-traitements sur Wikipédia pour augmenter la quantité d'exemples annotés, tout en introduisant une quantité raisonnable de bruit. Ensuite, nous explorons différentes méthodes d'utilisation de données obtenues par supervision distante pour l'apprentissage des représentations, principalement pour apprendre des représentations de mots classiques (statistiques) et contextuelles. À cause de sa position centrale pour de nombreuses applications du TALN, nous choisissons la reconnaissance d'entité nommée (NER) comme tâche principale. Nous expérimentons avec des bancs d’essai NER standards et nous observons des performances état de l’art. Ce faisant, nous étudions un cadre plus intéressant, à savoir l'amélioration des performances inter-domaines (généralisation). / Recent years have seen a leap in deep learning techniques that greatly changed the way Natural Language Processing (NLP) tasks are tackled. In a couple of years, neural networks and word embeddings quickly became central components to be adopted in the domain. Distant supervision (DS) is a well-used technique in NLP to produce labeled data from partially annotated examples. Traditionally, it was mainly used as training data in the absence of manual annotations, or as additional training data to improve generalization performances. In this thesis, we study how distant supervision can be employed within a modern deep learning based NLP framework. As deep learning algorithms gets better when massive amount of data is provided (especially for representation learning), we revisit the task of generating distant supervision data from Wikipedia. We apply post-processing treatments on the original dump to further increase the quantity of labeled examples, while introducing a reasonable amount of noise. Then, we explore different methods for using distant supervision data for representation learning, mainly to learn classic and contextualized word representations. Due to its importance as a basic component in many NLP applications, we choose Named-Entity Recognition (NER) as our main task. We experiment on standard NER benchmarks showing state-of-the-art performances. By doing so, we investigate a more interesting setting, that is, improving the cross-domain (generalization) performances.

Studien syftar till att få kunskap om hur laborativt material och arbetssätt framhålls och behandlas inom området bråk i olika läromedel. Vidare undersöker vi även hur lärare undervisar laborativt utifrån dessa läromedel. För att samla in data har vi använt oss av metodtriangulering i form av granskning av läromedel och intervjuer. De läromedel som har granskats är Favorit matematik, Eldorado och Prima matematik. För att bredda kunskapen inom området intervjuades lärare som var aktiva användare av dessa läromedlen i sin undervisning. Med hjälp av variationsteorin har vi även analyserat de laborativa inslagen i läromedlen för att se på dess relevans i relation till variationsmönstrena kontrastering och generalisering och därmed hur elevers lärande främjas genom läromedlens föreslagna arbetssätt. Resultatet visar att de tre granskade läromedlen skiljer sig åt samtidigt som de har många likheter. Samtliga läromedel har tagit i beaktning att det laborativa materialet behöver ta plats i undervisningen på ett eller annat sätt. Dessa läromedel ger eleverna möjlighet att utforska med något slags laborativt material. Studiens slutsats blir således att vi kan se att läromedlen, på ett eller annat sätt, bidrar med laborativt material och arbetssätt. Utifrån variationsteorins sätt att se på lärande passar Eldorados laborativa arbetssätt bäst in på de variationsmönster vi granskat. / The study aims to gain knowledge about how laboratory material and working methods are highlighted and treated in the field of fraction in different teaching materials, and how teachers describe that they teach based on this. In order to collect data, we have used method triangulation in the form of examination of teaching material and interviews. The teaching materials that have been examined are Favorite Mathematics, Eldorado and Prima Mathematics. To broaden the knowledge in the field, teacher 's who were active users of these teaching materials were interviewed in their teaching. Using the theory of variation, we have also analyzed the laboratory elements in the teaching materials to look at its relevance in relation to the variation patterns of contrast and generalization and thus how pupils' learning is promoted through the proposed teaching methods. The result shows that the three study materials examined differ at the same time as they have many similarities. All teaching materials have taken into account that the laboratory material needs to take place in the teaching in one way or another. These teaching materials allow students to explore with some kind of laboratory material. The study's conclusion thus becomes that we can see that the teaching aids, in one way or another, contribute with laboratory material and working methods. Based on variation theory's approach to learning, Eldorado's laboratory approach fits best with the variation patterns we examined.

L’objectif principal de cette thèse est de souligner que le compromis biais-variance n’est pas toujours vrai (p. ex. dans les réseaux neuronaux). Nous plaidons pour que ce manque d’universalité soit reconnu dans les manuels scolaires et enseigné dans les cours d’introduction qui couvrent le compromis. Nous passons d’abord en revue l’historique du compromis entre les biais et les variances, sa prévalence dans les manuels scolaires et certaines des principales affirmations faites au sujet du compromis entre les biais et les variances. Au moyen d’expériences et d’analyses approfondies, nous montrons qu’il n’y a pas de compromis entre la variance et le biais dans les réseaux de neurones lorsque la largeur du réseau augmente. Nos conclusions semblent contredire les affirmations de l’oeuvre historique de Geman et al. (1992). Motivés par cette contradiction, nous revisitons les mesures expérimentales dans Geman et al. (1992). Nous discutons du fait qu’il n’y a jamais eu de preuves solides d’un compromis dans les réseaux neuronaux lorsque le nombre de paramètres variait. Nous observons un phénomène similaire au-delà de l’apprentissage supervisé, avec un ensemble d’expériences d’apprentissage de renforcement profond. Nous soutenons que les révisions des manuels et des cours magistraux ont pour but de transmettre cette compréhension moderne nuancée de l’arbitrage entre les biais et les variances. / The main goal of this thesis is to point out that the bias-variance tradeoff is not always true (e.g. in neural networks). We advocate for this lack of universality to be acknowledged in textbooks and taught in introductory courses that cover the tradeoff. We first review the history of the bias-variance tradeoff, its prevalence in textbooks, and some of the main claims made about the bias-variance tradeoff. Through extensive experiments and analysis, we show a lack of a bias-variance tradeoff in neural networks when increasing network width. Our findings seem to contradict the claims of the landmark work by Geman et al. (1992). Motivated by this contradiction, we revisit the experimental measurements in Geman et al. (1992). We discuss that there was never strong evidence for a tradeoff in neural networks when varying the number of parameters. We observe a similar phenomenon beyond supervised learning, with a set of deep reinforcement learning experiments. We argue that textbook and lecture revisions are in order to convey this nuanced modern understanding of the bias-variance tradeoff.

In this essay some fundamental results from the theory of machine learning and neural networks are presented, with the goal of finally discussing bounds on the generalization error of neural networks, via Rademacher complexity. / I denna uppsats presenteras några grundläggande resultat från teorin kring maskininlärning och neurala nätverk, med målet att slutligen diskutera övre begräsningar på generaliseringsfelet hos neurala nätverk, via Rademachers komplexitet.

<p>The capacity of deep reinforcement learning policy networks has been found to affect the performance of trained agents. It has been observed that policy networks with more parameters have better training performance and generalization ability than smaller networks. In this work, we find cases where this does not hold true. We observe unimodal variance in the zero-shot test return of varying width policies, which accompanies a drop in both train and test return. Empirically, we demonstrate mostly monotonically increasing performance or mostly optimal performance as the width of deep policy networks increase, except near the variance mode. Finally, we find a scenario where larger networks have increasing performance up to a point, then decreasing performance. We hypothesize that these observations align with the theory of double descent in supervised learning, although with specific differences.</p>

<p>In this work, we study a class of Markovian information theoretic optimization problems motivated by the recent interests in incorporating mutual information as performance metrics which gives evident success in representation learning, feature extraction and clustering problems. In particular, we focus on the information bottleneck (IB) and privacy funnel (PF) methods and their recent multi-view, multi-source generalizations that gain attention because the performance significantly improved with multi-view, multi-source data. Nonetheless, the generalized problems challenge existing IB and PF solves in terms of the complexity and their abilities to tackle large-scale data. </p> <p>To address this, we study both the IB and PF under a unified framework and propose solving it through splitting methods, including renowned algorithms such as alternating directional method of multiplier (ADMM), Peaceman-Rachford splitting (PRS) and Douglas-Rachford splitting (DRS) as special cases. Our convergence analysis and the locally linear rate of convergence results give rise to new splitting method based IB and PF solvers that can be easily generalized to multi-view IB, multi-source PF. We implement the proposed methods with gradient descent and empirically evaluate the new solvers in both synthetic and real-world datasets. Our numerical results demonstrate improved performance over the state-of-the-art approach with significant reduction in complexity. Furthermore, we consider the practical scenario where there is distribution mismatch between training and testing data generating processes under a known bounded divergence constraint. In analyzing the generalization error, we develop new techniques inspired by the input-output mutual information approach and tighten the existing generalization error bounds.</p>

Tous les enfants entre quatre et sept ans sont susceptibles de commettre des erreurs de confusion de caractères réversibles, dont le correspondant en miroir correspond à un autre caractère du système d'écriture (par exemple, b, d, p, et q dans l'alphabet latin). Ces erreurs sont le produit de la généralisation en miroir, une propriété du système visuel qui nous permet initialement de reconnaître un visage, un animal, un objet indépendamment de son orientation latérale. Elle s'applique automatiquement (elle est dite heuristique) lors de la lecture et de l'écriture. Pour apprendre à lire et à écrire, nous mobilisons en effet une partie des réseaux de neurones initialement dévolus au traitement visuel des visages, des animaux, et des objets, grâce à un processus de plasticité cérébrale nommé recyclage neuronal. La lecture et l'écriture héritent alors de la propriété de généralisation en miroir, bien que celle-ci soit non-pertinente pour la reconnaissance des lettres réversibles b, d, p, et q. Vers sept ans, la fréquence des erreurs de confusion des lettres réversibles chute brusquement. Le principal objectif de cette thèse est de tester l'hypothèse selon laquelle la généralisation en miroir n'est pas entièrement « désapprise » comme précédemment suggéré, mais activement inhibée. Pour cela, nous avons conçu des paradigmes d'amorçage négatif que nous avons proposés à des adultes lettrés et à des enfants d'école primaire dans quatre études empiriques. Une cinquième étude, théorique, propose une nouvelle loi des apprentissages d'objets culturels récents (langage écrit, mathématiques) sur la base du double processus de « recyclage neuronal + contrôle inhibiteur ». / Every child aged four to seven is likely to commit confusion errors on reversible characters, whose mirror-image counterpart is another character in the writing system (for instance, b, d, p, and q in the Latin alphabet). These errors are produced by the mirror generalization process, a property of the visual system that initially allows us to recognize a face, an animal, or an object independently of the perceived profile. It automatically applies (it is said to be heuristic) to reading and writing. We actually mobilize part of the neuronal networks initially allotted to the visual processing of faces, animals, and objects for learning to read and write, thanks to a process of brain plasticity called neuronal recycling. Thus, reading and writing inherits the mirror generalization property, although it is irrelevant for the recognition of reversible letters b, d, p, and q. Around seven years of age, the frequency of errors of reversible letters confusions suddenly drops. The main objective of the present thesis is to test the hypothesis that the mirror generalization process is not entirely "unlearned" as hypothesized by previous studies but rather actively inhibited. To this aim, we designed negative priming paradigms that we proposed to literate adults and primary school children in four empirical studies. A fifth study, more theoretical, proposes a new law of learning recent cultural objects (written language, mathematics), based on the dual process of "neuronal recycling + inhibitory control".