Dissertations / Theses on the topic 'Machine theory of collective intelligence'

The idea of a United Nations (UN) intelligence system is widely misunderstood and debates about it seem to be both misplaced and anecdotal. The lack of a consistent theory on intelligence has fostered the widely held view that such a system is not feasible or incompatible with the UN collective security system. This dissertation takes as its central thesis, the question, of whether an intelligence system is both desirable and feasible within the UN structure. In spite of the fact that no known study has so far engaged with the subject matter at the depth presented in this dissertation, the study advances the concept of collective intelligence, and its implications for managing international conflicts. The dissertation examines existing barriers in efforts to interface intelligence system with the UN structure, and proposes that, with suitable refinements, the concept of intelligence need not be incompatible with the UN system. It is also argued that these constraints should not preclude evolutionary changes to include an intelligence system that is compatible with an organisation such as the UN. By developing a concept of collective intelligence, the thesis proposes theoretical frameworks that suggest a potential nature of a viable intelligence capability within the UN. The analysis is developed normatively and conceptually, which lead to a further conclusion that the UN already possesses an intelligence capability which exists in manner that is not recognised. The lack of scholarly efforts to ground such a system on a reasonable framework creates a vacuum in the study of international organisations, and in particular the United Nations system. At a minimum, this dissertation fills this gap.

This thesis investigates the rhetorical features of blogs that lend them dialogic strength as an online genre through the lens of Mikhail Bakhtin's theories of speech genres, utterances, and dialogism. As a relatively new online genre, blogs stem from previous genres (in print and online as well as verbal), but their emergence as a popular form of expression in our current culture demands attention to how blogs also offer us different rhetorical opportunities to meet our changing social exigencies as online subjects in the 21st century. This thesis was inspired by questions about how blogs redefine the rhetorical situation to alter our textual roles as readers, writers, and respondents in the new generic circumstances we encounter--and reproduce--online. Applying the framework of Henry Jenkins' Convergence Culture and Pierre Levy's Collective Intelligence, this thesis analyzes how blogs enable us as online subjects to add our utterances to our textual collective intelligence, which benefits from our personal experience and the epistemic conversations of blogs as online texts. In addition, it is also an inquiry into how the rhetorical circumstances of blogs as textual sites of collective intelligence can create a reciprocal learning environment in the writing classroom. I ultimately examine blogs through the lenses of alternative pedagogy--informed by David Wallace and Helen Rothschild Ewald's Mutuality in the Rhetoric and Composition Classroom and Xin Liu Gale's Teachers, Discourses, and Authority in the Postmodern Composition Classroom--to suggest the potential consequences of a writing education that includes how we are currently writing--and being written by--our culture's online generic practice of blogs.
M.A.
Department of English
Arts and Humanities
English MA

This thesis consists of four chapters. Chapter 1 focuses on theoretical results on high-order laplacian-based regularization in function estimation. We studied the iterated laplacian regularization in the context of supervised learning in order to achieve both nice theoretical properties (like thin-plate splines) and good performance over complex region (like soap film smoother). In Chapter 2, we propose an innovative static path-planning algorithm called m-A* within an environment full of obstacles. Theoretically we show that m-A* reduces the number of vertex. In the simulation study, our approach outperforms A* armed with standard L1 heuristic and stronger ones such as True-Distance heuristics (TDH), yielding faster query time, adequate usage of memory and reasonable preprocessing time. Chapter 3 proposes m-LPA* algorithm which extends the m-A* algorithm in the context of dynamic path-planning and achieves better performance compared to the benchmark: lifelong planning A* (LPA*) in terms of robustness and worst-case computational complexity. Employing the same beamlet graphical structure as m-A*, m-LPA* encodes the information of the environment in a hierarchical, multiscale fashion, and therefore it produces a more robust dynamic path-planning algorithm. Chapter 4 focuses on an approach for the prediction of spot electricity spikes via a combination of boosting and wavelet analysis. Extensive numerical experiments show that our approach improved the prediction accuracy compared to those results of support vector machine, thanks to the fact that the gradient boosting trees method inherits the good properties of decision trees such as robustness to the irrelevant covariates, fast computational capability and good interpretation.

Recent studies in cognitive science indicate that language has an important social function. The structure and knowledge of language emerges from the processes of human communication together with the domain-general cognitive processes. Each individual of a community interacts socially with a limited number of peers. Nevertheless societies are characterized by their stunning global regularities. By dealing with the language as a complex adaptive system, we are able to analyze how languages change and evolve over time. Multi-agent computational simulations assist scientists from different disciplines to build several language emergence scenarios. In this thesis several simulations are implemented and tested in order to categorize examples in a test data set efficiently and accurately by using a population of agents interacting by playing categorization games inspired by L. Steels'
s naming game. The emergence of categories throughout interactions between a population of agents in the categorization games are analyzed. The test results of categorization games as a model combination algorithm with various machine learning algorithms on different data sets have shown that categorization games can have a comparable performance with fast convergence.

The dissertation addresses the research topics of machine learning outlined below. We developed the theory about traditional first-order algorithms from convex opti- mization and provide new insights in nonconvex objective functions from machine learning. Based on the theory analysis, we designed and developed new algorithms to overcome the difficulty of nonconvex objective and to accelerate the speed to obtain the desired result. In this thesis, we answer the two questions: (1) How to design a step size for gradient descent with random initialization? (2) Can we accelerate the current convex optimization algorithms and improve them into nonconvex objective? For application, we apply the optimization algorithms in sparse subspace clustering. A new algorithm, CoCoSSC, is proposed to improve the current sample complexity under the condition of the existence of noise and missing entries. Gradient-based optimization methods have been increasingly modeled and inter- preted by ordinary differential equations (ODEs). Existing ODEs in the literature are, however, inadequate to distinguish between two fundamentally different meth- ods, Nesterov’s acceleration gradient method for strongly convex functions (NAG-SC) and Polyak’s heavy-ball method. In this paper, we derive high-resolution ODEs as more accurate surrogates for the two methods in addition to Nesterov’s acceleration gradient method for general convex functions (NAG-C), respectively. These novel ODEs can be integrated into a general framework that allows for a fine-grained anal- ysis of the discrete optimization algorithms through translating properties of the amenable ODEs into those of their discrete counterparts. As a first application of this framework, we identify the effect of a term referred to as gradient correction in NAG-SC but not in the heavy-ball method, shedding deep insight into why the for- mer achieves acceleration while the latter does not. Moreover, in this high-resolution ODE framework, NAG-C is shown to boost the squared gradient norm minimization at the inverse cubic rate, which is the sharpest known rate concerning NAG-C itself. Finally, by modifying the high-resolution ODE of NAG-C, we obtain a family of new optimization methods that are shown to maintain the accelerated convergence rates as NAG-C for minimizing convex functions.

This thesis considers the classification of physical states in a simplified gearbox using acoustical data and simple time domain signal shape characterisation techniques allied to a basic feedforward multi-layer perceptron neural network. A novel extension to the signal coding scheme (TES), involving the application of energy based shape descriptors, was developed. This sought specifically to improve the techniques suitability to the identification of mechanical states and was evaluated against the more traditional minima based TES descriptors. The application of learning based identification techniques offers potential advantages over more traditional programmed techniques both in terms of greater noise immunity and in the reduced requirement for highly skilled operators. The practical advantages accrued by using these networks are studied together with some of the problems associated in their use within safety critical monitoring systems.Practical trials were used as a means of developing the TES conversion mechanism and were used to evaluate the requirements of the neural networks being used to classify the data. These assessed the effects upon performance of the acquisition and digital signal processing phases as well as the subsequent training requirements of networks used for accurate condition classification. Both random data selection and more operator intensive performance based selection processes were evaluated for training. Some rudimentary studies were performed on the internal architectural configuration of the neural networks in order to quantify its influence on the classification process, specifically its effect upon fault resolution enhancement.The techniques have proved to be successful in separating several unique physical states without the necessity for complex state definitions to be identified in advance. Both the computational demands and the practical constraints arising from the use of these techniques fall within the bounds of a realisable system.

De nombreux algorithmes en Apprentissage Automatique utilisent une notion de distance ou de similarité entre les exemples pour résoudre divers problèmes tels que la classification, le partitionnement ou l'adaptation de domaine. En fonction des tâches considérées ces métriques devraient avoir des propriétés différentes mais les choisir manuellement peut-être fastidieux et difficile. Une solution naturelle est alors d'adapter automatiquement ces métriques à la tâche considérée. Il s'agit alors d'un problème connu sous le nom d'Apprentissage de Métriques et où le but est principalement de trouver les meilleurs paramètres d'une métrique respectant des contraintes spécifiques. Les approches classiques dans ce domaine se focalisent habituellement sur l'apprentissage de distances de Mahalanobis ou de similarités bilinéaires et l'une des principales limitations est le fait que le contrôle du comportement de ces métriques est souvent limité. De plus, si des travaux théoriques existent pour justifier de la capacité de généralisation des modèles appris, la plupart des approches ne présentent pas de telles garanties. Dans cette thèse nous proposons de nouveaux algorithmes pour apprendre des métriques à comportement contrôlé et nous mettons l'accent sur les propriétés théoriques de ceux-ci. Nous proposons quatre contributions distinctes qui peuvent être séparées en deux parties: (i) contrôler la métrique apprise en utilisant une métrique de référence et (ii) contrôler la transformation induite par la métrique apprise. Notre première contribution est une approche locale d'apprentissage de métriques où le but est de régresser une distance proportionnelle à la perception humaine des couleurs. Notre approche est justifiée théoriquement par des garanties en généralisation sur les métriques apprises. Dans notre deuxième contribution nous nous sommes intéressés à l'analyse théorique de l'intérêt d'utiliser une métrique de référence dans un terme de régularisation biaisé pour aider lors du processus d'apprentissage. Nous proposons d'utiliser trois cadres théoriques différents qui nous permettent de dériver trois mesures différentes de l'apport de la métrique de référence. Ces mesures nous donnent un aperçu de l'impact de la métrique de référence sur celle apprise. Dans notre troisième contribution nous proposons un algorithme d'apprentissage de métriques où la transformation induite est contrôlée. L'idée est que, plutôt que d'utiliser des contraintes de similarité et de dissimilarité, chaque exemple est associé à un point virtuel qui appartient déjà à l'espace induit par la métrique apprise. D'un point de vue théorique nous montrons que les métriques apprises de cette façon généralisent bien mais aussi que notre approche est liée à une méthode plus classique d'apprentissage de métriques basée sur des contraintes de paires. Dans notre quatrième contribution nous essayons aussi de contrôler la transformation induite par une métrique apprise. Cependant, plutôt que considérer un contrôle individuel pour chaque exemple, nous proposons une approche plus globale en forçant la transformation à suivre une transformation géométrique associée à un problème de transport optimal. D'un point de vue théorique nous proposons une discussion sur le lien entre la transformation associée à la métrique apprise et la transformation associée au problème de transport optimal. D'un point de vue plus pratique nous montrons l'intérêt de notre approche pour l'adaptation de domaine mais aussi pour l'édition d'images
Many Machine Learning algorithms make use of a notion of distance or similarity between examples to solve various problems such as classification, clustering or domain adaptation. Depending on the tasks considered these metrics should have different properties but manually choosing an adapted comparison function can be tedious and difficult. A natural trend is then to automatically tailor such metrics to the task at hand. This is known as Metric Learning and the goal is mainly to find the best parameters of a metric under some specific constraints. Standard approaches in this field usually focus on learning Mahalanobis distances or Bilinear similarities and one of the main limitations is that the control over the behaviour of the learned metrics is often limited. Furthermore if some theoretical works exist to justify the generalization ability of the learned models, most of the approaches do not come with such guarantees. In this thesis we propose new algorithms to learn metrics with a controlled behaviour and we put a particular emphasis on the theoretical properties of these algorithms. We propose four distinct contributions which can be separated in two parts, namely (i) controlling the metric with respect to a reference metric and (ii) controlling the underlying transformation corresponding to the learned metric. Our first contribution is a local metric learning method where the goal is to regress a distance proportional to the human perception of colors. Our approach is backed up by theoretical guarantees on the generalization ability of the learned metrics. In our second contribution we are interested in theoretically studying the interest of using a reference metric in a biased regularization term to help during the learning process. We propose to use three different theoretical frameworks allowing us to derive three different measures of goodness for the reference metric. These measures give us some insights on the impact of the reference metric on the learned one. In our third contribution we propose a metric learning algorithm where the underlying transformation is controlled. The idea is that instead of using similarity and dissimilarity constraints we associate each learning example to a so-called virtual point belonging to the output space associated with the learned metric. We theoretically show that metrics learned in this way generalize well but also that our approach is linked to a classic metric learning method based on pairs constraints. In our fourth contribution we also try to control the underlying transformation of a learned metric. However instead of considering a point-wise control we consider a global one by forcing the transformation to follow the geometrical transformation associated to an optimal transport problem. From a theoretical standpoint we propose a discussion on the link between the transformation associated with the learned metric and the transformation associated with the optimal transport problem. On a more practical side we show the interest of our approach for domain adaptation but also for a task of seamless copy in images

The project investigates mathematics, informatics, statistical analysis and their histories, the history of human engagement with machines, and illustrates some uses of artificial intelligence and robotic technologies as media. It is concerned with, amongst other issues, the sentient and not sentient binaries offered in discourses on machine intelligence. The term intelligence is used to distinguish between human and not human. However, a non-human, the intelligent machine, has become incorporated into the processes by which our culture defines intelligence. Those processes were explored in phases of the project that focused upon various kinds of interactions between people and machines, particularly the ways in which those interactions are mediated by knowledge. The discourses that underpin the field of mechanical intelligence spring from the same sources as the rhetoric that delineates human beings from all other things. We make intelligent machines because we have something to prove regarding our own intelligence. The devices expose attributes considered in our culture to be intelligent. The size and technical sophistication of modern robots result from the expenditure of considerable funds across several disciplines. Such machines signify wealth, power and excess, despite any other significance their makers intend.

Nowadays, probabilistic models are playing a significant role in various areas in- cluding machine learning, artificial intelligence and cognitive science, etc. How- ever, as those models are becoming more and more complex, it shows that the corresponding programs are really hard to maintain and reuse as well. Meanwhile, the current tools are not feasible enough to enable probabilistic modeling and ma- chine learning to be accessible to the working programmer, who has sufficient do- main expertise, but perhaps not enough expertise in probability theory or machine learning. Probabilistic programming is one possible way to solve this. Indeed, probabilistic programming languages are powerful tools to specify probabilistic models directly in terms of a computer programs. While programmers writes normal procedures, everything will be automatically translated into statistical distributions and then users can do inferences upon them. This project aims at exploring and implementing a probabilistic programming language, for which we name as Shelang. We use Scheme, a dialect of Lisp lan- guage which is originated from λ-Calculus, to implement a embedded probabilis- tic programming language. This paper mainly discusses about the design, algo- rithms, details of this implementation and several usages of Shelang and make a conclusion in the end.

In many settings, agents engage in problem-solving activities, which require them to share resources, act on each others behalf, coordinate individual acts, etc. If autonomous agents are to e ectively interact (or support interaction among humans) in situations such as deciding whom and how to approach the provision of a resource or the performance of an action, there are a number of important questions to address. Who do I choose to delegate a task to? What do I need to say to convince him/her to do something? Were similar requests granted from similar agents in similar circumstances? What arguments were most persuasive? What are the costs involved in putting certain arguments forward? Research in argumentation strategies has received signi cant attention in recent years, and a number of approaches has been proposed to enable agents to reason about arguments to present in order to persuade another. However, current approaches do not adequately address situations where agents may be operating under social constraints (e.g., policies) that regulate behaviour in a society. In this thesis, we propose a novel combination of techniques that takes into consideration the policies that others may be operating with. First, we present an approach where evidence derived from dialogue is utilised to learn the policies of others. We show that this approach enables agents to build more accurate and stable models of others more rapidly. Secondly, we present an agent decision-making mechanism where models of others are used to guide future argumentation strategy. This approach takes into account the learned policy constraints of others, the cost of revealing in- formation, and anticipated resource availability in deciding whom to approach. We empirically evaluate our approach within a simulated multi-agent frame- work, and demonstrate that through the use of informed strategies agents can improve their performance.

La problematique generale qui oriente notre travail concerne l'aptitude de groupes d'agents artificiels a expliquer leur comportement individuel et collectif. Cette recherche se situe au carrefour de l'intelligence artificielle distribuee et de la thematique explicative. Le propre d'un raisonnement effectue dans un systeme multi-agent est l'intervention de plusieurs agents autonomes et cooperatifs, qui participent, selon leur competence, a une activite collective finalisee. La construction des explications est fortement dependante des specificites du systeme (architecture, interaction, role et nature des agents). La demarche que nous proposons concerne des societes ou agents communiquent par envoi de messages (communication explicite, directe et volontaire). L'explication du comportement collectif du systeme doit tenir compte du double aspect de la definition d'un agent (son comportement interne et son role social) et peut etre construite selon deux niveaux d'abstraction (l'agent et/ou la societe). La juxtaposition d'explications individuelles etant insuffisante pour realiser une explication synthetique coherente, il est necessaire d'integrer dans la societe un agent explicatif. Il decide de la ligne d'explication, elabore puis assemble les explications individuelles. L'agent en cours d'explication est choisi relativement aux interactions dont il est l'origine ou le destinataire. Une telle demarche a ete implementee dans le cadre du systeme multi-agent synergic. La construction des explications est de la competence d'un agent particulier: l'agent explicatif. Depositaire d'un savoir expliquer independant de tout domaine d'application, il construit un texte lisible, pertinent et homogene malgre les changements d'agent etudie. Une analyse des resultats obtenus par l'eee sont presentes. Cette etude a permis d'envisager de nouvelles perspectives sur les roles d'un agent explicatif dans un systeme multi-agent

Curiosity of human nature drives us to explore the origins of what makes each of us different. From ancient legends and mythology, Mendel's law, Punnett square to modern genetic research, we carry on this old but eternal question. Thanks to technological revolution, today's scientists try to answer this question using easily measurable gene expression and other profiling data. However, the exploration can easily get lost in the data of growing volume, dimension, noise and complexity. This dissertation is aimed at developing new machine learning methods that take data from different classes as input, augment them with knowledge of feature relationships, and train classification models that serve two goals: 1) class prediction for previously unseen samples; 2) knowledge discovery of the underlying causes of class differences. Application of our methods in genetic studies can help scientist take advantage of existing biological networks, generate diagnosis with higher accuracy, and discover the driver networks behind the differences. We proposed three new graph-based regularization algorithms. Graph Connectivity Constrained AdaBoost algorithm combines a connectivity module, a deletion function, and a model retraining procedure with the AdaBoost classifier. Graph-regularized Linear Programming Support Vector Machine integrates penalty term based on submodular graph cut function into linear classifier's objective function. Proximal Graph LogisticBoost adds lasso and graph-based penalties into logistic risk function of an ensemble classifier. Results of tests of our models on simulated biological datasets show that the proposed methods are able to produce accurate, sparse classifiers, and can help discover true genetic differences between phenotypes.

Artificial intelligence has in recent years taken center stage in the technological development. Major corporations, operating in a variety of economic sectors, are investing heavily in AI in order to stay competitive in the years and decades to come. What differentiates this technology from traditional computing is that it can carry out tasks previously limited to humans. As such it contains the possibility to revolutionize every aspect of our society. Until now, social science has not given the proper attention that this emerging technological phenomena deserves, a phenomena which, according to some, is increasing in strength exponentially. This paper aims to problematize AI in the light of democratic elections, both as an analytical tool and as a tool for manipulation. It also looks at three recent empirical cases where AI technology was used extensively. The results show that there in fact are reasons to worry. AI as an instrument can be used to covertly affect the public debate, to depress voter turnout, to polarize the population, and to hinder understanding of political issues.

Thesis (M.S.)--Worcester Polytechnic Institute.
Title from title screen. Keywords: genetic programming; approximations; machine learning; artificial intelligence. Includes bibliographical references (p. 92-94).

Traditional views of grammatical theory hold that languages are characterised by sets of constructions. This approach entails the enumeration of all possible constructions for each language being described. Current theories of transformational generative grammar have established an alternative position. Specifically, Chomsky's Government-Binding theory proposes a system of principles which are common to human language. Such a theory is referred to as a "Universal Grammar"(UG). Associated with the principles of grammar are parameters of variation which account for the diversity of human languages. The grammar for a particular language is known as a "Core Grammar", and is characterised by an appropriately parametrised instance of UG. Despite these advances in linguistic theory, construction-based approaches have remained the status quo within the field of natural language processing. This thesis investigates the possibility of developing a principle-based system which reflects the modular nature of the linguistic theory. That is, rather than stipulating the possible constructions of a language, a system is developed which uses the principles of grammar and language specific parameters to parse language. Specifically, a system-is presented which performs syntactic analysis and translation for a subset of English and German. The cross-linguistic nature of the theory is reflected by the system which can be considered a procedural model of UG.
Science, Faculty of
Computer Science, Department of
Graduate

Most law firms have done little to address the opportunities and threats related to potentially disruptive technology (DT), such as artificial intelligence (AI) and machine learning (ML). The purpose of this multiple case study was to explore strategies that law firm leaders in the United States used to address the potentially detrimental influences of DT, such as AI and ML, on their organizations. The systems approach to management was employed as the conceptual framework. Data were collected from 6 participants at 2 international law firms with offices in California using semistructured interviews and organizational artifacts. Data were analyzed using Miles, Huberman, and Saldana's data analysis method, resulting in 4 themes: recognizing the legal ecosystem and legal firms are open systems, but organizational subsystems often function as semiclosed systems; acknowledging that while DT represents the most significant potential challenge in the near future, the immediate challenge is improving technology, which requires organizational adjustments; recognizing the need for firms to invest more heavily in innovation generation activities; and realizing the need for increased utilization of augmenting technologies, such as AI or ML, to streamline nonadvisory outputs. The findings of this study might support best practices for addressing DT and contribute to social change by outlining ways in which firms can lower costs to clients while increasing access to legal services for those in underserved communities.

Nous décrivons le phénomène d’engagement à travers des expériences transmédia coconstruites avec les bénéficiaires dans le contexte de la formation pour adultes. Nous abordons ce terrain en faisant l’hypothèse qu’un transmédia à valeur expérientielle permet d’accroitre la participation des sujets pour consolider l’engagement en formation. Cette hypothèse ouvre la perspective d’une médiation de type transmédia capable, d’intégrer des objectifs d’exploitation scientifique de l’engagement et de valorisation de la participation et de l’attention qui pourraient être intéressants pour d’autres corpus.À partir d’un cadre d’analyse centré sur les bénéficiaires des transactions, notre enquête met en exergue différentes formes d’engagement suivant des transmédia hybrides aux caractéristiques spécifiques. La complémentarité de ces transmédia a favorisé divers régimes d’engagement observés de façon ponctuelle, pour un engagement sur la durée
We describe the phenomenon of engagement through co-constructed transmedia experiences with the beneficiaries in the context of adult education. We approach this ground on the assumption that a transmedia with experiential value makes it possible to increase the participation of subjects to consolidate the commitment in training. This hypothesis opens up the prospect of a transmedia type of mediation capable of integrating the objectives of scientific exploitation for the commitment and enhancement of participation and the attention that could be interesting for other corpuses. Based on an analysis framework focused on the beneficiaries of transactions, our survey highlighted different forms of hybrid transmedia engagement with specific characteristics. The complementarity of these transmedia has favoured various commitment regimes observed on an ad hoc basis for a long-term commitment

Recently, strategies of National Basketball Association teams have evolved with the skillsets of players and the emergence of advanced analytics. One of the most effective actions in dynamic offensive strategies in basketball is the dribble hand-off (DHO). This thesis proposes an architecture for a classification pipeline for detecting DHOs in an accurate and automated manner. This pipeline consists of a combination of player tracking data and event labels, a rule set to identify candidate actions, manually reviewing game recordings to label the candidates, and embedding player trajectories into hexbin cell paths before passing the completed training set to the classification models. This resulting training set is examined using the information gain from extracted and engineered features and the effectiveness of various machine learning algorithms. Finally, we provide a comprehensive accuracy evaluation of the classification models to compare various machine learning algorithms and highlight their subtle differences in this problem domain.

L'Apprentissage Automatique tire ses racines d'un large champ disciplinaire qui inclut l'Intelligence Artificielle, la Reconnaissance de Formes, les Statistiques ou l'Optimisation. Dès les origines de l'Apprentissage, les questions computationelles et les propriétés en généralisation ont toutes deux été identifiées comme centrales pour la discipline. Tandis que les premières concernent les questions de calculabilité ou de complexité (sur un plan fondamental) ou d'efficacité computationelle (d'un point de vue plus pratique) des systèmes d'apprentissage, les secondes visent a comprendre et caractériser comment les solutions qu'elles fournissent vont se comporter sur de nouvelles données non encore vues. Ces dernières années, l'émergence de jeux de données à grande échelle en Apprentissage Automatique a profondément remanié les principes de la Théorie de l'Apprentissage. En prenant en compte de potentielles contraintes sur le temps d'entraînement, il faut faire face à un compromis plus complexe que ceux qui sont classiquement traités par les Statistiques. Une conséquence directe tient en ce que la mise en place d'algorithmes efficaces (autant en théorie qu'en pratique) capables de tourner sur des jeux de données a grande échelle doivent impérativement prendre en compte les aspects statistiques et computationels de l'Apprentissage de façon conjointe. Cette thèse a pour but de mettre à jour, analyser et exploiter certaines des connections qui existent naturellement entre les aspects statistiques et computationels de l'Apprentissage. Plus précisément, dans une première partie, nous étendons l'analyse en stabilité, qui relie certaines propriétés algorithmiques aux capacités de généralisation des algorithmes d'apprentissage, la matrice de confusion, que nous suggérons comme nouvelle mesure de performance (fine). Dans une seconde partie, nous présentons un nouvelle approche pour apprendre une fonction de régression basée sur les noyaux, où le noyau appris sert directement la tâche de régression, et qui exploite la structure du problème pour offrir une procédure d'optimisation peu coûteuse. Finalement, nous étudions le compromis entre vitesse de convergence et coût computationel lorsque l'on minimise une fonction composite avec des méthodes par gradient-proximal inexact. Dans ce contexte, nous identifions des stratégies d'optimisation qui sont computationellement optimales.

Orientador: Jacques Wainer
Tese (doutorado) - Universidade Estadual de Campinas, Instituto de Computação
Made available in DSpace on 2018-08-09T21:11:41Z (GMT). No. of bitstreams: 1 JambeiroFilho_JorgeEduardodeSchoucair_D.pdf: 736285 bytes, checksum: b7d7f186f743f9b0e541c857b0ca8226 (MD5) Previous issue date: 2007
Resumo: Analisamos o uso de métodos Bayesianos em um problema de classificação de padrões de interesse prático para a Receita Federal do Brasil que é caracterizado pela presença de atributos de alta cardinalidade e pela existência de interações relevantes entre eles. Mostramos que a presença de atributos de alta cardinalidade pode facilmente gerar tantas subdivisões no conjunto de treinamento que, mesmo tendo originalmente uma grande quantidade de dados, acabemos obtendo probabilidades pouco confiáveis, inferidas a partir de poucos exemplos. Revisamos as estratégias usualmente adotadas para lidar com esse problema dentro do universo Bayesiano, exibindo sua dependência em suposições de não interação inaceitáveis em nosso domínio alvo. Mostramos empiricamente que estratégias Bayesianas mais avançadas para tratamento de atributos de alta cardinalidade, como pré-processamento para redução de cardinalidade e substituição de tabelas de probabilidades condicionais (CPTs) de redes Bayesianas (BNs) por tabelas default (DFs), árvores de decisão (DTs) e grafos de decisão (DGs) embora tragam benefícios pontuais não resultam em ganho de desempenho geral em nosso domínio alvo. Propomos um novo método Bayesiano de classificação, chamado de hierarchical pattern Bayes (HPB), que calcula probabilidades posteriores para as classes dado um padrão W combinando as observações de W no conjunto de treinamento com probabilidades prévias que são obtidas recursivamente a partir das observações de padrões estritamente mais genéricos que W. Com esta estratégia, ele consegue capturar interações entre atributos de alta cardinalidade quando há dados suficientes para tal, sem gerar probabilidades pouco confiáveis quando isso não ocorre. Mostramos empiricamente que, em nosso domínio alvo, o HPB traz benefícios significativos com relação a redes Bayesianas com estruturas populares como o naïve Bayes e o tree augmented naïve Bayes, com relação a redes Bayesianas (BNs) onde as tabelas de probabilidades condicionais foram substituídas pelo noisy-OR, por DFs, por DTs e por DGs, e com relação a BNs construídas, após uma fase de redução de cardinalidade usando o agglomerative information bottleneck. Além disso, explicamos como o HPB, pode substituir CPTs e mostramos com testes em outro problema de interesse prático que esta substituição pode trazer ganhos significativos. Por fim, com testes em vários conjuntos de dados públicos da UCI, mostramos que a utilidade do HPB ser bastante ampla
Abstract: In this work, we analyze the use of Bayesian methods in a pattern classification problem of practical interest for Brazil¿s Federal Revenue which is characterized by the presence of high cardinality attributes and by the existence of relevant interactions among them.We show that the presence of high cardinality attributes can easily produce so many subdivisions in the training set that, even having originally a great amount of data, we end up with unreliable probability estimates, inferred from small samples. We cover the most common strategies to deal with this problem within the Bayesian universe and show that they rely strongly on non interaction assumptions that are unacceptable in our target domain. We show empirically that more advanced strategies to handle high cardinality attributes like cardinality reduction by preprocessing and conditional probability tables replacement with default tables, decision trees and decision graphs, in spite of some restricted benefits, do not improve overall performance in our target domain. We propose a new Bayesian classification method, named hierarchical pattern Bayes (HPB), which calculates posterior class probabilities given a pattern W combining the observations of W in the training set with prior class probabilities that are obtained recursively from the observations of patterns that are strictly more generic than W. This way, it can capture interactions among high cardinality attributes when there is enough data, without producing unreliable probabilities when there is not. We show empirically that, in our target domain, HPB achieves significant performance improvements over Bayesian networks with popular structures like naïve Bayes and tree augmented naïve Bayes, over Bayesian networks where traditional conditional probability tables were substituted by noisy-OR gates, default tables, decision trees and decision graphs, and over Bayesian networks constructed after a cardinality reduction preprocessing phase using the agglomerative information bottleneck method. Moreover, we explain how HPB can replace conditional probability tables of Bayesian Networks and show, with tests in another practical problem, that such replacement can result in significant benefits. At last, with tests over several UCI datasets we show that HPB may have a quite wide applicability
Doutorado
Sistemas de Informação
Doutor em Ciência da Computação

Dans le domaine de l'apprentissage machine, les caractéristiques des données varient généralement dans l'espace des entrées : la distribution globale pourrait être multimodale et contenir des non-linéarités. Afin d'obtenir de bonnes performances, l'algorithme d'apprentissage devrait alors être capable de capturer et de s'adapter à ces changements. Même si les modèles linéaires ne parviennent pas à décrire des distributions complexes, ils sont réputés pour leur passage à l'échelle, en entraînement et en test, aux grands ensembles de données en termes de nombre d'exemples et de nombre de fonctionnalités. Plusieurs méthodes ont été proposées pour tirer parti du passage à l'échelle et de la simplicité des hypothèses linéaires afin de construire des modèles aux grandes capacités discriminatoires. Ces méthodes améliorent les modèles linéaires, dans le sens où elles renforcent leur expressivité grâce à différentes techniques. Cette thèse porte sur l'amélioration des approches d'apprentissage locales, une famille de techniques qui infère des modèles en capturant les caractéristiques locales de l'espace dans lequel les observations sont intégrées.L'hypothèse fondatrice de ces techniques est que le modèle appris doit se comporter de manière cohérente sur des exemples qui sont proches, ce qui implique que ses résultats doivent aussi changer de façon continue dans l'espace des entrées. La localité peut être définie sur la base de critères spatiaux (par exemple, la proximité en fonction d'une métrique choisie) ou d'autres relations fournies, telles que l'association à la même catégorie d'exemples ou un attribut commun. On sait que les approches locales d'apprentissage sont efficaces pour capturer des distributions complexes de données, évitant de recourir à la sélection d'un modèle spécifique pour la tâche. Cependant, les techniques de pointe souffrent de trois inconvénients majeurs :ils mémorisent facilement l'ensemble d'entraînement, ce qui se traduit par des performances médiocres sur de nouvelles données ; leurs prédictions manquent de continuité dans des endroits particuliers de l'espace ; elles évoluent mal avec la taille des ensembles des données. Les contributions de cette thèse examinent les problèmes susmentionnés dans deux directions : nous proposons d'introduire des informations secondaires dans la formulation du problème pour renforcer la continuité de la prédiction et atténuer le phénomène de la mémorisation ; nous fournissons une nouvelle représentation de l'ensemble de données qui tient compte de ses spécificités locales et améliore son évolutivité. Des études approfondies sont menées pour mettre en évidence l'efficacité de ces contributions pour confirmer le bien-fondé de leurs intuitions. Nous étudions empiriquement les performances des méthodes proposées tant sur des jeux de données synthétiques que sur des tâches réelles, en termes de précision et de temps d'exécution, et les comparons aux résultats de l'état de l'art. Nous analysons également nos approches d'un point de vue théorique, en étudiant leurs complexités de calcul et de mémoire et en dérivant des bornes de généralisation serrées
In Machine Learning field, data characteristics usually vary over the space: the overall distribution might be multi-modal and contain non-linearities.In order to achieve good performance, the learning algorithm should then be able to capture and adapt to these changes. Even though linear models fail to describe complex distributions, they are renowned for their scalability, at training and at testing, to datasets big in terms of number of examples and of number of features. Several methods have been proposed to take advantage of the scalability and the simplicity of linear hypotheses to build models with great discriminatory capabilities. These methods empower linear models, in the sense that they enhance their expressive power through different techniques. This dissertation focuses on enhancing local learning approaches, a family of techniques that infers models by capturing the local characteristics of the space in which the observations are embedded. The founding assumption of these techniques is that the learned model should behave consistently on examples that are close, implying that its results should also change smoothly over the space. The locality can be defined on spatial criteria (e.g. closeness according to a selected metric) or other provided relations, such as the association to the same category of examples or a shared attribute. Local learning approaches are known to be effective in capturing complex distributions of the data, avoiding to resort to selecting a model specific for the task. However, state of the art techniques suffer from three major drawbacks: they easily memorize the training set, resulting in poor performance on unseen data; their predictions lack of smoothness in particular locations of the space;they scale poorly with the size of the datasets. The contributions of this dissertation investigate the aforementioned pitfalls in two directions: we propose to introduce side information in the problem formulation to enforce smoothness in prediction and attenuate the memorization phenomenon; we provide a new representation for the dataset which takes into account its local specificities and improves scalability. Thorough studies are conducted to highlight the effectiveness of the said contributions which confirmed the soundness of their intuitions. We empirically study the performance of the proposed methods both on toy and real tasks, in terms of accuracy and execution time, and compare it to state of the art results. We also analyze our approaches from a theoretical standpoint, by studying their computational and memory complexities and by deriving tight generalization bounds

Les robots vont peu à peu intégrer nos foyers sous la forme d’assistants et de compagnons,humanoïdes ou non. Afin de remplir leur rôle efficacement ils devront s’adapter àl’utilisateur, notamment en apprenant de celui-ci le savoir ou les capacités qui leur fontdéfaut. Dans ce but, leur manière d’interagir doit être naturelle et évoquer les mêmesmécanismes coopératifs que ceux présent chez l’homme. Au centre de ces mécanisme setrouve le concept d’action : qu’est-ce qu’une action, comment les humains les reconnaissent,comment les produire ou les décrire ? La modélisation de toutes ces fonctionnalitésconstituera la fondation de cette thèse et permettra la mise en place de mécanismescoopératifs de plus haut niveau, en particulier les plan partagés qui permettent à plusieursindividus d’oeuvrer de concert afin d’atteindre un but commun. Finalement, je présenteraiune différence fondamentale entre la représentation de la connaissance chez l’homme etchez la machine, toujours dans le cadre de l’interaction coopérative : la dissociation possibleentre le corps d’un robot et sa cognition, ce qui n’est pas imaginable chez l’homme. Cettedissociation m’amènera notamment à explorer le « shared experience framework », unesituation dans laquelle une cognition artificielle centrale gère l’expérience partagée demultiples individus ayant chacun une identité propre. Cela m’amènera finalement àquestionner les différentes philosophies de l’esprit du point de vue de l’attribution d’unesprit à une machine et de ce que cela impliquerai quant à l’esprit humain
Robots will gradually integrate our homes wielding the role of companions, humanoids ornot. In order to cope with this status they will have to adapt to the user, especially bylearning knowledge or skills from him that they may lack. In this context, their interactionshould be natural and evoke the same cooperative mechanisms that humans use. At thecore of those mechanisms is the concept of action: what is an action, how do humansrecognize them, how they produce or describe them? The modeling of aspects of thesefunctionalities will be the basis of this thesis and will allow the implementation of higherlevel cooperative mechanisms. One of these is the ability to handle “shared plans” whichallow two (or more) individuals to cooperate in order to reach a goal shared by all.Throughout the thesis I will attempt to make links between the human development ofthese capabilities, their neurophysiology, and their robotic implementation. As a result ofthis work, I will present a fundamental difference between the representation of knowledgein humans and machines, still in the framework of cooperative interaction: the possibledissociation of a robot body and its cognition, which is not easily imaginable for humans.This dissociation will lead me to explore the “shared experience framework, a situationwhere a central artificial cognition manages the shared knowledge of multiple beings, eachof them owning some kind of individuality. In the end this phenomenon will interrogate thevarious philosophies of mind by asking the question of the attribution of a mind to amachine and the consequences of such a possibility regarding the human mind

In this work, I view incremental experiential learning in intelligent software agents as progressive agent self-adaptation. When an agent produces an incorrect behavior, then it may reflect on, and thus diagnose and repair, the reasoning and knowledge that produced the incorrect behavior. In particular, I focus on the self-diagnosis and self-repair of an agent's domain knowledge. The implementation of systems with the capability to self-diagnose and self-repair involves building both reasoning processes capable of such learning and knowledge representations capable of supporting those reasoning processes. The core issue my dissertation addresses is: what kind of metaknowledge (knowledge about knowledge) may enable the agent to diagnose faults in its domain knowledge? In providing a solution to this issue, the central contribution of this research is a theory of the kind of metaknowledge that enables a system to reason about and adapt its conceptual knowledge. For this purpose, I propose a representation that explicitly encodes metaknowledge in the form of procedures called Empirical Verification Procedures (EVPs). In the proposed knowledge representation, an EVP is associated with each concept within the agent's domain knowledge. Each EVP explicitly semantically grounds the associated concept in the agent's perception, and can thus be used as a test to determine the validity of knowledge of that concept during diagnosis. I present the formal and empirical evaluation of a system, Augur, that makes use of EVP metaknowledge to adapt its own domain knowledge in the context of a particular subclass of classification problem that I call compositional classification, in which the overall classification task can be broken into a hierarchically organized set of subtasks. I hypothesize that EVP metaknowledge will enable a system to automatically adapt its knowledge in two ways: first, by adjusting the ways that inputs are categorized by a concept, in accordance with semantics fixed by an associated EVP; and second, by adjusting the semantics of concepts themselves when they fail to contribute appropriately to system goals. The latter adaptation is realized by altering the EVP associated with the concept in question. I further hypothesize that the semantic grounding of domain concepts in perception through the use of EVPs will increase the generalization power of a learner that operates over those concepts, and thus make learning more efficient. Beyond the support of these hypotheses, I also present results pertinent to the understanding of learning in compositional classification settings using structured knowledge representations.

Orientador: Fernando José Von Zuben
Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Engenharia Elétrica e de Computação
Made available in DSpace on 2018-08-20T00:09:41Z (GMT). No. of bitstreams: 1 Arruda_RodrigoLopesSettide_M.pdf: 2295891 bytes, checksum: 4f5f4bc8f219b0c3c27239520027d496 (MD5) Previous issue date: 2012
Resumo: Com o uso de sistemas cognitivos em uma crescente gama de aplicações, criou-se uma grande expectativa e elevada demanda por máquinas cada vez mais autônomas, inteligentes e criativas na solução de problemas reais. Em diversos casos, os desafios demandam capacidade de aprendizado e adaptação. Este trabalho lida com conceitos de aprendizagem por reforço e discorre sobre as principais abordagens de solução e variações de problemas. Em seguida, constrói uma proposta híbrida incorporando outras ideias em aprendizagem de máquina, validando-a com experimentos simulados. Os experimentos permitem apontar as principais vantagens da metodologia proposta, a qual está fundamentada em sua capacidade de lidar com cenários de espaços contínuos e, também, de aprender uma política ótima enquanto segue outra, exploratória. A arquitetura proposta é híbrida, baseada em uma rede neural perceptron multi-camadas acoplada a um aproximador de funções denominado wirefitting. Esta arquitetura é coordenada por um algoritmo adaptativo e dinâmico que une conceitos de programação dinâmica, análise de Monte Carlo, aprendizado por diferença temporal e elegibilidade. O modelo proposto é utilizado para resolver problemas de controle ótimo, por meio de aprendizagem por reforço, em cenários com variáveis contínuas e desenvolvimento não-linear. Duas instâncias diferentes de problemas de controle, reconhecidas na literatura pertinente, são apresentadas e testadas com a mesma arquitetura
Abstract: With the evergrowing use of cognitive systems in various applications, it has been created a high expectation and a large demand for machines more and more autonomous, intelligent and creative in real world problem solving. In several cases, the challenges ask for high adaptive and learning capability. This work deals with the concepts of reinforcement learning, and reasons on the main solution approaches and problem variations. Subsequently, it builds a hybrid proposal incorporating other machine learning ideas, so that the proposal is validated with simulated experiments. The experiments allow to point out the main advantages of the proposed methodology, founded on its capability to handle continuous space environments, and also to learn an optimal policy while following an exploratory policy. The proposed architecture is hybrid in the sense that it is based on a multi-layer perceptron neural network coupled with a function approximator called wire-fitting. The referred architecture is coordinated by a dynamic and adaptive algorithm which merges concepts from dynamic programming, Monte Carlo analysis, temporal difference learning, and eligibility. The proposed model is used to solve optimal control problems, by means of reinforcement learning, in scenarios endowed with continuous variables and nonlinear development. Two different instances of control problems, well discussed in the pertinent literature, are presented and tested with the same architecture
Mestrado
Engenharia de Computação
Mestre em Engenharia Elétrica

Due to the long duration required to perform manual knowledge entry by human knowledge engineers it is desirable to find methods to automatically acquire knowledge about the world by accessing online information. In this work I examine using the Cyc ontology to guide the creation of Naïve Bayes classifiers to provide knowledge about items described in Wikipedia articles. Given an initial set of Wikipedia articles the system uses the ontology to create positive and negative training sets for the classifiers in each category. The order in which classifiers are generated and used to test articles is also guided by the ontology. The research conducted shows that a system can be created that utilizes statistical text classification methods to extract information from an ad-hoc generated information source like Wikipedia for use in a formal semantic ontology like Cyc. Benefits and limitations of the system are discussed along with future work.

The present study shows that prestructuring based on domain knowledge leads to statistically significant generalization-performance improvement in artificial neural networks (NNs) of the multilayer perceptron (MLP) type, specifically in the case of a noisy real-world problem with numerous interacting variables. The prestructuring of MLPs based on knowledge of the structure of a problem domain has previously been shown to improve generalization performance. However, the problem domains for those demonstrations suffered from significant shortcomings: 1) They were purely logical problems, and 2) they contained small numbers of variables in comparison to most data-mining applications today. Two implications of the former were a) the underlying structure of the problem was completely known to the network designer by virtue of having been conceived for the problem at hand, and b) noise was not a significant concern in contrast with real-world conditions. As for the size of the problem, neither computational resources nor mathematical modeling techniques were advanced enough to handle complex relationships among more than a few variables until recently, so such problems were left out of the mainstream of prestructuring investigations. In the present work, domain knowledge is built into the solution through Reconstructability Analysis, a form of information-theoretic modeling, which is used to identify mathematical models that can be transformed into a graphic representation of the problem domain's underlying structure. Employing the latter as a pattern allows the researcher to prestructure the MLP, for instance, by disallowing certain connections in the network. Prestructuring reduces the set of all possible maps (SAPM) that are realizable by the NN. The reduced SAPM--according to the Lendaris-Stanley conjecture, conditional probability, and Occam's razor--enables better generalization performance than with a fully connected MLP that has learned the same I/O mapping to the same extent. In addition to showing statistically significant improvement over the generalization performance of fully connected networks, the prestructured networks in the present study also compared favorably to both the performance of qualified human agents and the generalization rates in classification through Reconstructability Analysis alone, which serves as the alternative algorithm for comparison.

This thesis investigates the design of sociable technologies and is divided into three main parts described below. In the first part, we introduce sociable technologies. We review our the definition of technology and propose categories of technologies according to the motivation underlying their design: improvement of control, improvement of communication or improvement of cooperation. Sociable technologies are then presented as an extension of techniques to improve cooperation. The design of sociable technologies are then discussed leading to the observation that acquisition of social common sense is a key challenge for designing sociable technologies. Finally, polite technologies are presented as an approach for acquiring social common sense. In the second part, we focus on the premises for the design of sociable technologies. A key aspect of social common sense is the ability to act appropriately in social situations. Associating appropriate behaviour with social situations is presented as a key method for implementing polite technologies. Reinforcement learning is proposed as a method for learning such associations and variation of this algorithm are experimentally evaluated. Learning the association between situation and behaviour relies on the strong assumption that mutual understanding of social situations can be achieved between technologies and people during interaction. We argue that in order to design sociable technologies, we must change the model of communication used by our technologies. We propose to replace the well-known code model of communication, with the ostensive-inferential model proposed by Sperber and Wilson. Hypotheses raised by this approach are evaluated in an experiment conducted in a smart environment, where, subjects by group of two or three are asked to collaborate with a smart environment in order to teach it how to behave in an automated meeting. A novel experimental methodology is presented: The Sorceress of Oz. The results collected from this experiment validate our hypothesis and provide insightful information for the design. We conclude by presenting, what we believe are, the premises for the design of sociable technologies. The final part of the thesis concerns an infrastructure for the design of sociable technologies. This infrastructure provides the support for three fundamental components. First, it provides the support for an inferential model of context. This inferential model of context is presented; a software architecture is proposed and evaluated in an experiment conducted in a smart-environment. Second, it provides the support for reasoning by analogy and introduces the concept of eigensituations. The advantage of this representation are discussed and evaluated in an experiment. Finally, it provides the support for ostensive-inferential communication and introduces the concept of ostensive interface
This thesis investigates the design of sociable technologies and is divided into three main parts described below. In the first part, we introduce sociable technologies. We review our the definition of technology and propose categories of technologies according to the motivation underlying their design: improvement of control, improvement of communication or improvement of cooperation. Sociable technologies are then presented as an extension of techniques to improve cooperation. The design of sociable technologies are then discussed leading to the observation that acquisition of social common sense is a key challenge for designing sociable technologies. Finally, polite technologies are presented as an approach for acquiring social common sense. In the second part, we focus on the premises for the design of sociable technologies. A key aspect of social common sense is the ability to act appropriately in social situations. Associating appropriate behaviour with social situations is presented as a key method for implementing polite technologies. Reinforcement learning is proposed as a method for learning such associations and variation of this algorithm are experimentally evaluated. Learning the association between situation and behaviour relies on the strong assumption that mutual understanding of social situations can be achieved between technologies and people during interaction. We argue that in order to design sociable technologies, we must change the model of communication used by our technologies. We propose to replace the well-known code model of communication, with the ostensive-inferential model proposed by Sperber and Wilson. Hypotheses raised by this approach are evaluated in an experiment conducted in a smart environment, where, subjects by group of two or three are asked to collaborate with a smart environment in order to teach it how to behave in an automated meeting. A novel experimental methodology is presented: The Sorceress of Oz. The results collected from this experiment validate our hypothesis and provide insightful information for the design. We conclude by presenting, what we believe are, the premises for the design of sociable technologies. The final part of the thesis concerns an infrastructure for the design of sociable technologies. This infrastructure provides the support for three fundamental components. First, it provides the support for an inferential model of context. This inferential model of context is presented; a software architecture is proposed and evaluated in an experiment conducted in a smart-environment. Second, it provides the support for reasoning by analogy and introduces the concept of eigensituations. The advantage of this representation are discussed and evaluated in an experiment. Finally, it provides the support for ostensive-inferential communication and introduces the concept of ostensive interface

La collaboration entre startups et grandes entreprises s’avère de plus en plus nécessaire dans le contexte actuel d’innovation ouverte, d’accélération de la demande du marché et donc de course de plus en plus rapide à l’innovation. Ces partenaires asymétriques présentent toutefois des différences notoires qui peuvent générer une distance entre eux, susceptible de mettre en péril le projet de collaboration. Au-delà de la dyade, d’autres acteurs de l’écosystème, en particulier les intermédiaires de l’innovation, participent également au projet de collaboration. L’objectif de cette thèse est de faire émerger les facteurs favorisant la collaboration symbiotique entre startups et grandes entreprises, basés sur une indépendance organisationnelle et financière des acteurs. Cette thèse vise également à montrer l’intérêt du recours à l’analogie avec la symbiose biologique entre symbiotes interagissant dans un écosystème donné. Il s’agit ainsi de mettre en exergue les facteurs d’équilibre de la relation, dans une optique gagnant-gagnant. Partant des différences mises en lumière à travers la distance cognitive, cette recherche propose d’étudier le phénomène de la collaboration startup - grande entreprise selon une approche exploratoire et une méthode mixte, qualitative et quantitative, basée sur la méthode des cas. L’étude de 38 cas réalisée (ayant conduit à une collecte de données auprès de 53 répondants sous forme d’entretiens et de questionnaire) propose une approche temporelle, multi-perspectives et holistique, mobilisant le cadre théorique de la proximité (géographique, cognitive, sociale, organisationnelle) et celui des capacités dynamiques. Cette recherche a donné lieu à quatre articles conduisant à plusieurs contributions théoriques et managériales. En premier lieu, l’étude selon la perspective des startups a permis d’identifier les facteurs favorisant la proximité et la collaboration entre startups et grandes entreprises en fonction de quatre niveaux : intra-organisationnel de la grande entreprise, intra-organisationnel de la startup, inter-organisationnel et écosystémique. La poursuite de l’exploration a mis en lumière la complémentarité des compétences des équipes fondatrices de startups, en comparaison avec les startupers solos, et qui est source de proximité avec les grandes entreprises. La suite de l’étude, selon la perspective des grandes entreprises, a mis en évidence l’importance d’un management fondé sur l’intelligence collective ainsi que de l’évolution du rôle des managers intermédiaires des grandes entreprises dans l’implémentation d’une stratégie d’innovation ouverte intégrant une variété d’acteurs, telles que les startups. Enfin, l’étude de la perspective des intermédiaires de l’innovation quant à leurs rôles dans le développement de la collaboration startup - grande entreprise a permis de faire émerger ces différents rôles selon trois phases de la construction de la collaboration, dont celui de constituer une ressource externe pour la grande entreprise pour la régénération de ses capacités dynamiques. Une contribution transversale est également l’identification et l’opérationnalisation des 2+1 phases de la collaboration suivant un axe chronologique : les phases Amont, de Design et de Processus de la collaboration
Collaboration between startups and large firms is becoming increasingly necessary in the current context of open innovation, accelerating market demand and thus the increasingly rapid race to innovate. These asymmetrical partners, however, present significant differences that can generate a distance between them that can jeopardize the collaboration project. Beyond the dyad, other actors of the ecosystem, in particular innovation intermediaries, also participate in the collaborative project. The objective of this thesis is to bring out the factors fostering symbiotic collaboration between startups and large firms, based on an organizational and financial independence of the actors. This thesis also aims to show the interest of using the analogy with the biological symbiosis between symbionts interacting in a given ecosystem. The aim is thus to highlight the balance factors of the relationship, in a win-win perspective. Starting from the differences brought to light through cognitive distance, this research proposes to study the phenomenon of startup - large firm collaboration according to an exploratory approach and a mixed qualitative and quantitative method, based on the case method. The study of 38 cases carried out (leading to a data collection from 53 respondents in the form of interviews and survey) proposes a time-based, multi-perspective and holistic approach, mobilizing the theoretical framework of proximity (geographical, cognitive, social, organizational) and that of dynamic capabilities. This research resulted in four articles leading to several theoretical and managerial contributions. Firstly, the study from the startup's perspective allowed to identify the factors fostering proximity and collaboration between startups and large firms according to four levels: intra-organizational of the large firm, intra-organizational of the startup, inter-organizational and ecosystemic. Further exploration has then highlighted the complementary skills of startup founding teams, compared to solo startuppers, which is a source of proximity to large firms. The continuation of the study, from the perspective of large firms, brought to light the importance of a management based on collective intelligence as well as the evolving role of middle managers in large firms in the implementation of an open innovation strategy integrating a variety of actors, such as startups. Finally, the study of the perspective of innovation intermediaries regarding their roles in the development of startup - large firm collaboration has allowed these different roles to emerge according to three phases of the collaboration construction, including that of constituting an external resource for the large firm for the regeneration of its dynamic capabilities. A transversal contribution is also the identification and operationalization of the 2+1 phases of the collaboration along a chronological axis: the Upstream, Design and Process phases of the collaboration

In our current political climate, state level legislators have become increasingly impor- tant. Due to cuts in funding and growing focus at the national level, public oversight for these legislators has drastically decreased. This makes it difficult for citizens and activists to understand the relationships and commonalities between legislators. This thesis provides three contributions to address this issue. First, we created a data set containing over 1200 features focused on a legislator’s activity on bills. Second, we created embeddings that represented a legislator’s level of activity and engagement for a given bill using a custom model called Democracy2Vec. Third, we provided a case study focused on the 2015-2016 California State Legislator and had our results verified by a political expert. Our results show that our embeddings can explain relationships between legislator and how they will likely act during the legislative process.

Nous nous positionnons dans un contexte d'informatique ambiante dans lequel il arrive que les besoins de l'utilisateur n'aient pas été prévus, notamment en situation exceptionnelle. Dans ce cas, il peut ne pas exister de système préconçu qui réponde exactement à ces besoins. Pour les satisfaire, il faut alors pouvoir composer les systèmes disponibles dans l'environnement, et le système composé doit permettre à l'utilisateur de faire des choix à l'exécution. Ainsi, l'utilisateur a la possibilité d'adapter l'exécution de la composition à son contexte. Cela signifie que la composition intègre des structures de contrôle de l'exécution, destinées à l'utilisateur : la composition est dite flexible. Dans cette thèse, nous proposons de répondre au problème de la composition flexible en contexte d'intelligence ambiante avec un planificateur produisant des plans flexibles. Dans un premier temps, nous proposons une modélisation de la planification flexible. Pour cela, nous définissons les opérateurs de séquence et d'alternative, utilisés pour caractériser les plans flexibles. Nous définissons deux autres opérateurs au moyen de la séquence et de l'alternative : l'entrelacement et l'itération. Nous nous référons à ce cadre théorique pour délimiter la flexibilité traitée par notre planificateur Lambda-Graphplan. L'originalité de Lambda-Graphplan est de produire des itérations en s'appuyant sur une approche par graphe de planification. Nous montrons notamment que Lambda-Graphplan est très performant avec les domaines se prêtant à la construction de structures itératives.

The appearance of the world varies dramatically not only from place to place but also from hour to hour and month to month. Billions of images that capture this complex relationship are uploaded to social-media websites every day and often are associated with precise time and location metadata. This rich source of data can be beneficial to improve our understanding of the globe. In this work, we propose a general framework that uses these publicly available images for constructing dense maps of different ground-level attributes from overhead imagery. In particular, we use well-defined probabilistic models and a weakly-supervised, multi-task training strategy to provide an estimate of the expected visual and auditory ground-level attributes consisting of the type of scenes, objects, and sounds a person can experience at a location. Through a large-scale evaluation on real data, we show that our learned models can be used for applications including mapping, image localization, image retrieval, and metadata verification.

Made available in DSpace on 2015-05-14T12:36:26Z (GMT). No. of bitstreams: 1 parte1.pdf: 1614982 bytes, checksum: 20a76dbb04a57c702c64e525119a04ec (MD5) Previous issue date: 2011-02-07
Coordenação de Aperfeiçoamento de Pessoal de Nível Superior
Several technological advances during recent years provided that the Electronic Health Record systems (EHR) became a solidified and viable alternative to replace progres-sively and efficiently, the use of health records on paper. The benefits found are associated with the use of methods for clinical decision support (CDS), data availability, ease in finding information, among other advantages inherent in computerized systems use. However, there are still many challenges and research to get the full potential of such systems. For example, the amounts of clinical data for EHR storage are very high. Several interests might benefit if there was a tool capable of performing an automated analysis, or more commonly found, semi-automated, useful for search patterns in the data set stored in the system. Several studies indicate that efforts in the field of machine learning achieve great results in various areas including clinical information. However, the effort required is still high, increasing the time spent with planning and processing, with high costs and large amounts of data needed for processing. This work, in association with the OpenCTI's CDS seeks to significantly reduce the amount of effort necessary to promote both the reuse of clin-ical information from the automatic learning, and the development of mechanisms for clini-cal decision support with low cost. This study seeks to offer those benefits to users of EHR systems, through a simple mechanism, but extensive, for analysis of clinical data stored in clinical databases. This anal-ysis is performed using a methodology of knowledge extraction algorithms using collective intelligence or data mining, through steps of search, selection, preprocessing, modeling, evaluation and application of the information extracted from these systems. From this, me-chanisms for clinical decision support of EHR, may use the framework offered by CollectMed to promote with greater ease and precision, more accurate information regarding specific medical conditions on their patients, according to what has already been registered by health professionals in similar cases using the EHR.
Diversos avanços tecnológicos ocorridos nos últimos anos fizeram com que os Sis-temas de Registro Eletrônico em Saúde (RES) se consolidassem como uma alternativa viável para substituir, progressivamente e com eficiência, o uso dos registros de saúde em papel. Os benefícios encontrados são associados ao uso de métodos de apoio à decisão clínica, disponi-bilidade dos dados, facilidade na busca por informações, entre outras vantagens inerentes ao uso de sistemas computadorizados. Entretanto, existem ainda, muitos desafios e pesquisas para fazer com que todo o potencial desses sistemas seja utilizado. Por exemplo, a quantida-de de dados clínicos que os sistemas de RES armazenam, é muito elevado. Diversos interes-ses poderiam ser beneficiados, caso houvesse uma ferramenta capaz de realizar uma análise automatizada, ou semi-automatizada (como é mais comumente encontrada), para buscar padrões úteis no conjunto de dados armazenados no sistema. Diversos trabalhos apontam que os esforços realizados no campo de aprendizado automático alcançam ótimos resultados em diversas áreas, inclusive para informações clíni-cas. Porém, o esforço necessário ainda é elevado, aumentando o tempo dedicado ao planeja-mento e execução, assim como altos custos e necessidade de grande volume de dados para o processamento. Este trabalho, associado ao sistema de apoio à decisão do OpenCTI busca reduzir, significativamente,o esforço necessário para promover tanto o reuso de informações clínicas a partir do aprendizado automático, quanto o desenvolvimento de mecanismos de apoio à decisão clínica a um baixo custo. O presente trabalho, busca oferecer tal benefício aos usuários de sistemas de RES, por meio de um mecanismo simples, porém amplo, de análise dos dados clínicos armazena-dos nos bancos de dados dos RES. Essa análise será realizada por meio de uma metodologia de extração de conhecimento, utilizando algoritmos de inteligência coletiva ou data mining, passando por etapas de busca, seleção, pré-processamento, modelagem, avaliação e aplicação destas informações extraídas dos sistemas. A partir disso, mecanismos de apoio à decisão clínica dos RES, poderão utilizar o arcabouço oferecido pelo CollectMed para promover, com mais facilidade e precisão, recuperação de informações mais apuradas a respeito das condi-ções clínicas específicas sobre seus pacientes, de acordo com o que já foi registrado por pro-fissionais de saúde em casos clínicos semelhantes persistidos no RES.

This thesis explores the metaphoric cyborg concept, based on the classic essay "A Cyborg Manifesto" by Donna J. Haraway, through Claudia Springer and her analysis of cyborgs in popular culture, and in relation to Judith Butler and queer theory. By extending the cyborg term to include contemporary forms of cyborgship such as our internet lives and personas, I advocate discussion of philosophical matters concering man/machine-compounds today rather than tomorrow. The main purpose is to restore the cyborg concept as a meaningful analytic and philosophic tool for studying man/machine-relations, in a world where other theories concerning nature/culture, man/machine/animal, subject/object, and similar dichotomies, partly lack a technological perspective. Hopefully, this thesis manages to connect three decades of cyborg reflections from several points of view, posing interesting questions about our dealings and feelings towards our dear friends – the machines.

Orientador : Fernando Jose Von Zuben
Tese (doutorado) - Universidade Estadual de Campinas, Faculdade de Engenharia Eletrica e de Computação
Made available in DSpace on 2018-08-04T02:17:19Z (GMT). No. of bitstreams: 1 Lima_ClodoaldoAparecidodeMoraes_D.pdf: 5444612 bytes, checksum: 0172ca4143c2737bf19a3b2076c06b44 (MD5) Previous issue date: 2004
Resumo: Os algoritmos baseados em métodos de kernel destacam-se entre as diversas técnicas de aprendizado de máquina. Eles foram inicialmente empregados na implementação de máquinas de vetores-suporte (SVMs). A abordagem SVM representa um procedimento de aprendizado não-paramétrico para classificação e regressão de alto desempenho. No entanto, existem aspectos estruturais e paramétricos de projeto que podem conduzir a uma degradação de desempenho. Na ausência de uma metodologia sistemática e de baixo custo para a proposição de modelos computacionais otimamente especificados, os comitês de máquinas se apresentam como alternativas promissoras. Existem versões estáticas de comitês, na forma de ensembles de componentes, e versões dinâmicas, na forma de misturas de especialistas. Neste estudo, os componentes de um ensemble e os especialistas de uma mistura são tomados como SVMs. O objetivo é explorar conjuntamente potencialidades advindas de SVM e comitê de máquinas, adotando uma formulação unificada. Várias extensões e novas configurações de comitês de máquinas são propostas, com análises comparativas que indicam ganho significativo de desempenho frente a outras propostas de aprendizado de máquina comumente adotadas para classificação e regressão
Abstract: Algorithms based on kernel methods are prominent techniques among the available approaches for machine learning. They were initially applied to implement support vector machines (SVMs). The SVM approach represents a nonparametric learning procedure devoted to high performance classification and regression tasks. However, structural and parametric aspects of the design may guide to performance degradation. In the absence of a systematic and low-cost methodology for the proposition of optimally specified computational models, committee machines emerge as promising alternatives. There exist static versions of committees, in the form of ensembles of components, and dynamic versions, in the form of mixtures of experts. In the present investigation, the components of an ensemble and the experts of a mixture are taken as SVMs. The aim is to jointly explore the potentialities of both SVM and committee machine, by means of a unified formulation. Several extensions and new configurations of committee machines are proposed, with comparative analyses that indicate significant gain in performance before other proposals for machine learning commonly adopted for classification and regression
Doutorado
Engenharia de Computação
Doutor em Engenharia Elétrica

Today defence systems are becoming more complex as technology advances and it is of great importance to explore new ways of solving problems and keep national defence current. In particular, Artificial Intelligence (AI) is used in an increasing number of industries such as logistic solutions, inventory management and defence. This thesis will evaluate the possibility to use Reinforcement Learning (RL) in an Air Defence Coordination(ADC) scenario at Saab AB. To evaluate RL, a simplified ADC-scenario is considered and solved using two different methods, Q-learning and Deep Q-learning (DQL). The results of the two methods are discussed as well as the limitations in scope and complexity for Q-learning. Deep Q-learning, on the other hand shows to be relatively easy to apply to more complicated scenarios. Finally, one last experiment with a far more complex scenario is constructed in order to show the scalability of DQL and create a foundation for future work in this field.
Dagens försvarssystem blir allt mer komplexa när tekniken utvecklas och det blir allt viktigare att utforska nya sätt att lösa problem för att ha ett toppmodernt försvar. I synnerhet används Artificiell intelligens (AI) i ett ökande antal branscher så som logistik, lagerhantering och försvar. Detta arbete kommer att utvärdera möjligheten att använda Förstärkt inlärning (RL) i ett Koordinerat luftförsvar (ADC) scenario hos Saab AB. För att utvärdera RL, löses ett förenklat ADC-scenario med två olika metoder, Q-learning och Deep Q-learning (DQL). Resultatet av de två metoderna diskuteras så väl som begränsningar för Q-learning. Å andra sidan visar sig DQL vara relativt enkelt att tillämpa i ett mer komplext scenario. Slutligen görs ett sista experiment med ett mycket mer komplicerat scenario för att visa skalbarheten för DQL och skapa en naturlig övergång till framtida arbete.

Computer Science is currently facing a grand challenge :finding good design practices for embedded systems. Embedded systems are essentially computers interacting with some physical process. You could find one in a braking systems or in a nuclear power plant for example. They present several design difficulties :first they are reactive systems, interacting indefinitely with their environment. Second,they must satisfy real-time constraints specifying when they should respond, and not only how. Finally, their environment is often deeply continuous, presenting complex dynamics. The formal models of choice for specifying such systems are timed and hybrid automata for which model checking is pretty well studied.

In a first part of this thesis, we study a complete design approach, including verification and code generation, for timed automata. We have to define a new semantics for timed automata, the AASAP semantics, that preserves the decidability properties for model checking and at the same time is implementable. Our notion of implementability is completely novel, and relies on the simulation of a semantics that is obviously implementable on a real platform. We wrote tools for the analysis and code generation and exemplify them on a case study about the well known Philips Audio Control Protocol.

In a second part of this thesis, we study the problem of controller synthesis for an environment specified as a hybrid automaton. We give a new solution for discrete controllers having only an imperfect information about the state of the system. In the process, we defined a new algorithm, based on the monotonicity of the controllable predecessors operator, for efficiently finding a controller and we show some promising applications on a classical problem :the universality test for finite automata.
Doctorat en sciences, Spécialisation Informatique
info:eu-repo/semantics/nonPublished

Cette thèse porte sur la détection de maladies cardio-vasculaires grâce à la surveillance de signaux physiologiques. L’objectif est de développer des approches de classification mono- et multi-label, basées sur la théorie des fonctions de croyance, pour prédire ou diagnostiquer une complication liée à une ou plusieurs maladies cardio-vasculaires. D’abord, une approche proposant une extraction de paramètres et une modélisation de l'information dans un cadre évidentiel est développée pour prédire la fibrillation atriale, une arythmie cardiaque. Une extension de cette approche utilise une option de rejet de classification et une modélisation alternative de l'information. La thèse élargit par la suite le domaine d’application pour couvrir plusieurs maladies cardio-vasculaires à la fois. Le problème est alors défini comme une classification multi-label où les labels représentent des traits caractéristiques des maladies. Une approche de classification multi-label est développée dans le domaine évidentiel qui se sert des corrélations entre les maladies pour accroître la précision du diagnostic. Enfin, une approche théorique de classification multi-label qui profite de la corrélation entre les labels a été proposée. Cette méthode d’ensemble permet de classifier en multi-label de manière efficace. Les approches proposées sont validées grâce à une base de données médicale publique, MIMIC III, hébergée sur Physionet
This thesis focuses on the detection of cardiovascular diseases through the monitoring of physiological signals. The objective is to develop mono- and multi-label classification approaches, based on the theory of belief functions, to predict or diagnose a complication linked to one or more cardiovascular diseases. First, an approach providing parameter extraction and information modeling in an evidential framework is developed to predict atrial fibrillation, a cardiac arrhythmia. An extension of this approach uses a reject classification option and alternative information modeling. The thesis then broadens the field of application to cover several cardiovascular diseases at the same time. The problem is thus defined as a multi-label classification where the labels represent features of the diseases. A multi-label classification approach is developed in the evidential domain which makes use of correlations between diseases to increase diagnostic accuracy. Finally, a theoretical approach of multi-label classification, which takes advantage of the correlation between labels, has been proposed. This ensemble method allows for efficient multi-label classification. The proposed approaches are validated using a public medical database, MIMIC III, hosted on Physionet

A customers propensity to buy a certain product is a widely researched field and is applied in multiple industries. In this thesis it is showed that using binary classification on data from Scandinavian Airlines can predict their customers propensity to book a flight within the next coming seven days. A comparison between logistic regression and support vector machine is presented and logistic regression with reduced number of variables is chosen as the final model, due to it’s simplicity and accuracy. The explanatory variables contains exclusively booking history, whilst customer demographics and search history is showed to be insignificant.
En kunds benägenhet att göra ett visst köp är ett allmänt undersökt område som applicerats i flera olika branscher. I den här studien visas det att statistiska binära klassificeringsmodeller kan användas för att prediktera Scandinavian Airlines kunders benägenhet att köpa en resa de kommande sju dagarna. En jämförelse är presenterad mellan logistisk regression och stödvektormaskin och logistisk regression med reducerat antal parametrar väljs som den slutgiltiga modellen tack vare sin enkelhet och träffsäkerhet. De förklarande variablerna är uteslutande bokningshistorik medan kundens demografi och sökdata visas vara insignifikant.

O presente trabalho apresenta dois novas modelos conexionistas, baseados na teoria da adaptação ressonante (ART): Simplified Fuzzy ARTMAP e Semantic ART (SMART). Descreve-se a modelagem, adaptação, implementação e validação destes, enquanto incorporados ao sistema hibrido HYCONES, para resolução de problemas de diagnostico medico em cardiopatias congênitas e nefrologia. HYCONES é uma ferramenta para a construção de sistemas especialistas híbridos que integra redes neurais com frames, assimilando as qualidades inerentes aos dois paradigmas. 0 mecanismo de frames fornece tipos construtores flexíveis para a modelagem do conhecimento do domínio, enquanto as redes neurais, representadas na versão original de HYCONES pelo modelo neural combinatório (MNC), possibilitam tanto a automação da aquisição de conhecimento, a partir de uma base de casos, quanta a implementação de aprendizado indutivo e dedutivo. A teoria da adaptação ressonante 6 caracterizada, principalmente, pela manutenção do equilíbrio entre as propriedades de plasticidade e estabilidade durante o processo de aprendizagem. ART inclui vários modelos conexionistas, tais como: Fuzzy ARTMAP, Fuzzy ART, ART 1, ART 2 e ART 3. Dentre estes, a rede neural Fuzzy ARTMAP destaca-se por possibilitar o tratamento de padr6es analógicos a partir de dois módulos ART básicos. O modelo Simplified Fuzzy ARTMAP, como o pr6prio nome o diz, a uma simplificação da rede neural Fuzzy ARTMAP. Ao contrario desta, o novo modelo possibilita o tratamento de padrões analógicos, a partir de apenas um modulo ART, responsável pelo tratamento dos padrões de entrada, adicionado de uma camada, responsável pelos padrões alvo. Mesmo com apenas um modulo ART, o modelo Simplified Fuzzy ARTMAP 6 capaz de reter o mesmo nível de desempenho obtido com a rede neural Fuzzy ARTMAP pois, continua a garantir, conjuntamente, a maximização da generalização e a minimização do erro preditivo, através da execução da estratégia match-tracking. Para a construção da base de casos de cardiopatias congênitas, 66 prontuários médicos, das três cardiopatias congênitas mais freqüentes, foram extraídos do banco de dados de pacientes submetidos a cirurgia cardíaca no Instituto de Cardiologia RS (ICFUC-RS). Tais prontuários abrangem o período de janeiro de 1986 a dezembro de 1990 e reportam 22 casos de Comunicação Interatrial (CIA), 29 de Comunicação Interventricular (CIV) e 15 de Defeito Septal Atrioventricular (DSAV). Para a análise de desempenho do sistema, 33 casos adicionais, do referido período, foram extraídos aleatoriamente do banco de dados do ICFUC-RS. Destes 33 casos, 13 apresentam CIA, 10 CIV e 10 DSAV. Para a construção da base de casos de síndromes renais, 381 prontuários do banco de dados de síndromes renais da Escola Paulista de Medicina foram analisados e 58 evidencias, correspondentes a dados de hist6ria clinica e exame físico dos pacientes, foram extraídas semi-automaticamente. Do total de casos selecionados, 136 apresentam Uremia, 85 Nefrite, 100 Hipertensão e 60 Litiase. Dos 381 casos analisados, 254 foram escolhidos aleatoriamente para a composicao do conjunto de treinamento, enquanto que os demais foram utilizados para a elaboração do conjunto de testes. Para que HYCONES II fosse validado, foram construídas 46 versões da base de conhecimento hibrida (BCH) para o domínio de cardiopatias congênitas e 46 versões da BCH para o de nefrologia. Em ambos os domínios médicos as respectivas bases de conhecimento foram construídas, automaticamente, a partir das respectivas bases de casos de treinamento. Das 46 versões geradas para cada grupo, uma representa o modelo MNC e 45 os modelos ART. As versões ART dividem-se em grupos de 3: 15 versões foram formadas a partir do modelo Simplified Fuzzy ARTMAP; 15 a partir deste mesmo modelo, sem que os padrões de entrada fossem normalizados; e, finalmente, 15 para o modelo Semantic ART. Na base de testes CHD, o desempenho da versa° HYCONES II - Simplified Fuzzy ARTMAP foi semelhante ao da versa° MNC. A primeira acertou 29 dos 33 diagnósticos (87,9%), enquanto a segunda apontou corretamente 31 dos 33 diagnósticos apresentados (93,9%). Na base de testes de síndromes renais, o desempenho de HYCONES II Fuzzy ARTMAP foi superior ao da versão MNC (p < 0,05). Ambas -Simplified acertaram, respectivamente, 108 (85%) e 95 (74,8%) diagnósticos, em 127 casos submetidos. Ainda que o desempenho da versão HYCONES II - Simplified Fuzzy ARTMAP se revelasse promissor, ao se examinar o conteúdo das redes geradas por este modelo, pode-se observar que estas divergiam completamente daquelas obtidas pelo MNC. As redes que levaram a conclusão diagnostica, na versão HYCONES - MNC, possuíam conteúdo praticamente igual aos grafos de conhecimento, elicitados de especialistas em cardiopatias congênitas. JA, as redes ativadas na versa° HYCONES II - Simplified Fuzzy ARTMAP, além de representarem numero bem major de evidencias que as redes MNC, a grande maioria destas ultimas representam a negação do padrão de entrada. Este fato deve-se a um processo de normalização, inerente ao modelo Simplified Fuzzy ARTMAP, no qual cada padrão de entrada e duplicado. Nesta duplicação, são representadas as evidências presentes em cada caso e, ao mesmo tempo, complementarmente, as evidencias ausentes, em relação ao total geral das mesmas na base de casos. Esta codificação inviabiliza o mecanismo de explanação do sistema HYCONES, pois, na área módica, os diagnósticos costumam ser feitos a partir de um conjunto de evidencias presentes e, não, pela ausência delas. Tentou-se, então, melhorar o conteúdo semântico das redes Simplified Fuzzy ARTMAP. Para tal, o processo de normalização ou codificação complementar da implementação do modelo foi retirado, validando-o novamente, contra o mesma base de testes. Na base de testes CHD, o desempenho de HYCONES II - Simplified Fuzzy ARTMAP, sem a codificação complementar, foi inferior ao da versão MNC (p < 0,05). A primeira acertou 25 dos 33 diagnósticos (75,8%), enquanto a segunda apontou corretamente 31 dos mesmos (93,9%). Na base de testes renais, o desempenho da versa° HYCONES II - Simplified Fuzzy ARTMAP, sem a codificação complementar, foi semelhante ao da versa° MNC. Dos 127 casos apresentados, a primeira acertou 98 diagn6sticos (77,2%), contra 95 da segunda (74,8%). Constatou-se, ainda, que as categorias de reconhecimento formadas pelo modelo Simplified Fuzzy ARTMAP continuavam a apresentar diferenças marcantes quanto ao seu conteúdo, quando comparadas as redes MNC ou aos grafos de conhecimento elicitados de especialistas. O modelo Semantic ART foi, então, proposto, na tentativa de se melhorar o conteúdo semantic° das redes ART. Modificou-se, então, o algoritmo de aprendizado do modelo Simplified Fuzzy ARTMAP, introduzindo-se o mecanismo de aprendizado indutivo do modelo MNC, i.e., o algoritmo de punições e recompensas, associado ao de poda e normalização. Nova validação com a mesma base de testes foi realizada. Para a base de testes de CHD, o desempenho de HYCONES II - SMART foi semelhante ao da versão Simplified Fuzzy ARTMAP e da versão MNC. A primeira e a segunda acertaram 29 dos 33 diagnósticos (87,9%), enquanto a versão MNC apontou corretamente 31 dos 33 diagnósticos apresentados (93,9%). Na base de testes de síndromes renais, o desempenho de HYCONES II - SMART foi superior ao da versão MNC (p < 0,05) e igual ao da versão Simplified Fuzzy ARTMAP. A primeira e a Ultima acertaram 108 dos 127 diagnósticos (85%), enquanto a segunda apontou corretamente 95 dos mesmos (74,8%). Desta feita, observou-se que as redes neurais geradas por HYCONES II - SMART eram semelhantes em conteúdo as redes MNC e aos grafos de conhecimento elicitados de múltiplos especialistas. As principais contribuições desta dissertação são: o projeto, implementação e validação dos modelos Simplified Fuzzy ARTMAP e SMART. Destaca-se, porem, o modelo SMART, que apresentou major valor semântico nas categorias de reconhecimento do que o observado nos modelos ART convencionais, graças a incorporação dos conceitos de especificidade e relevância. Esta dissertação, entretanto, representa não só a modelagem e validação de dois novos modelos neurais, mas sim, o enriquecimento do sistema HYCONES, a partir da continuação de dissertação de mestrado previamente defendida. A partir do presente trabalho, portanto, é dada a possibilidade de escolha, ao engenheiro de conhecimento, de um entre três modelos neurais: o MNC, o Semantic ART e o Simplified Fuzzy ARTMAP que, sem exceção, apresentam Born desempenho. Os dois primeiros destacam-se, contudo, por suportarem semanticamente o contexto.
This dissertation presents two new connectionist models based on the adaptive resonance theory (ART): Simplified Fuzzy ARTMAP and Semantic ART (SMART). The modeling, adaptation, implementation and validation of these models are described, in their association to HYCONES, a hybrid connectionist expert system to solve classification problems. HYCONES integrates the knowledge representation mechanism of frames with neural networks, incorporating the inherent qualities of the two paradigms. While the frames mechanism provides flexible constructs for modeling the domain knowledge, neural networks, implemented in HYCONES' first version by the combinatorial neuron model (CNM), provide the means for automatic knowledge acquisition from a case database, enabling, as well, the implementation of deductive and inductive learning. The Adaptive Resonance Theory (ART) deals with a system involving selfstabilizing input patterns into recognition categories, while maintaining a balance between the properties of plasticity and stability. ART includes a series of different connectionist models: Fuzzy ARTMAP, Fuzzy ART, ART 1, ART 2, and ART 3. Among them, the Fuzzy ARTMAP one stands out for being capable of learning analogical patterns, using two basic ART modules. The Simplified Fuzzy ARTMAP model is a simplification of the Fuzzy ARTMAP neural network. Constrating the first model, the new one is capable of learning analogical patterns using only one ART module. This module is responsible for the categorization of the input patterns. However, it has one more layer, which is responsible for receiving and propagating the target patterns through the network. The presence of a single ART module does not hamper the Simplified Fuzzy ARTMAP model. The same performance levels are attained when the latter one runs without the second ART module. This is certified by the match-tracking strategy, that conjointly maximizes generalization and minimizes predictive error. Two medical domains were chosen to validate HYCONES performance: congenital heart diseases (CHD) and renal syndromes. To build up the CHD case base, 66 medical records were extracted from the cardiac surgery database of the Institute of Cardiology RS (ICFUC-RS). These records cover the period from January 1986 to December 1990 and describe 22 cases of Atrial Septal Defect (ASD), 29 of Ventriculal Septal Defect (VSD), and 15 of Atrial- Ventricular Septa! Defect (AVSD), the three most frequent congenital heart diseases. For validation purposes, 33 additional cases, from the same database and period mentioned above, were also extracted. From these cases, 13 report ASD, 10 VSD and 10 AVSD. To build the renal syndromes case base, 381 medical records from the database of the Escola Paulista de Medicina were analyzed and 58 evidences, covering the patients' clinical history and physical examination data, were semiautomatically extracted. From the total number of selected cases, 136 exhibit Uremia, 85 Nephritis, 100 Hypertension, and 60 Calculosis. From the 381 cases analyzed, 245 were randomically chosen to build the training set, while the remaining ones were used to build the testing set. To validate HYCONES II, 46 versions of the hybrid knowledge base (HKB) with congenital heart diseases were built; for the renal domain, another set of 46 HKB versions were constructed. For both medical domains, the HKBs were automatically generated from the training databases. From these 46 versions, one operates with the CNM model and the other 45 deals with two ART models. These ART versions are divided in three groups: 15 versions were built using the Simplified Fuzzy ARTMAP model; 15 used the Simplified Fuzzy ARTMAP model without the normalization of the input patterns, and 15 used the Semantic ART model. HYCONES II - Simplified Fuzzy ARTMAP and HYCONES - CNM performed similarly for the CH D domain. The first one pointed out correctly to 29 of the 33 testing cases (87,9%), while the second one indicated correctly 31 of the same cases (93,9%). In the renal syndromes domain, however, the performance of HYCONES II - Simplified Fuzzy ARTMAP was superior to the one exhibited by CNM (p < 0,05). Both versions pointed out correctly, respectively, 108 (85%) and 95 (74.8%) diagnoses of the 127 testing cases presented to the system. HYCONES II - Simplified Fuzzy ARTMAP, therefore, displayed a satisfactory performance. However, the semantic contents of the neural nets it generated were completely different from the ones stemming from the CNM version. The networks that pointed out the final diagnosis in HYCONES - CNM were very similar to the knowledge graphs elicited from experts in congenital heart diseases. On the other hand, the networks activated in HYCONES II - Simplified Fuzzy ARTMAP operated with far more evidences than the CNM version. Besides this quantitative difference, there was a striking qualitative discrepancy among these two models. The Simplified Fuzzy ARTMAP version, even though pointing out to the correct diagnoses, used evidences that represented the complementary coding of the input pattern. This coding, inherent to the Simplified Fuzzy ARTMAP model, duplicates the input pattern, generating a new one depicting the evidence observed (on-cell) and, at the same time, the absent evidence, in relation to the total evidence employed to represent the input cases (off-cell). This coding shuts out the HYCONES explanation mechanism, since medical doctors usually reach a diagnostic conclusion rather from a set of observed evidences than from their absence. The next step taken was to improve the semantic contents of the Simplified Fuzzy ARTMAP model. To achieve this, the complement coding process was removed and the modified model was, then, revalidated, through the same testing sets as above described. In the CHD domain, the performance of HYCONES II - Simplified Fuzzy ARTMAP, without complementary coding, proved to be inferior to the one presented by CNM (p < 0,05). The first model singled out correctly 25 out of the 33 testing cases (75,8%), while the second one singled out correctly 31 out of the same 33 cases (93,9%). In the renal syndromes domain, the performances of HYCONES II - Simplified Fuzzy ARTMAP, without complementary coding, and HYCONES - CNM were similar. The first pointed out correctly to 98 of the 127 testing cases (77,2%), while the second one pointed out correctly to 95 of the same cases (74.8%). However, the recognition categories formed by this modified Simplified Fuzzy ARTMAP still presented quantitative and qualitative differences in their contents, when compared to the networks activated by CNM and to the knowledge graphs elicited from experts. This discrepancy, although smaller than the one observed in the original Fuzzy ARTMAP model, still restrained HYCONES explanation mechanism. The Semantic ART model (SMART) was, then, proposed. Its goal was to improve the semantic contents of ART recognition categories. To build this new model, the Simplified Fuzzy ARTMAP archictecture was preserved, while its learning algorithm was replaced by the CNM inductive learning mechanism (the punishments and rewards algorithm, associated with the pruning and normalization mechanisms). A new validation phase was, then, performed over the same testing sets. For the CHD domain, the perfomance comparison among SMART, Simplified Fuzzy ARTMAP, and CNM versions showed similar results. The first and the second versions pointed out correctly to 29 of the 33 testing cases (87,9%), while the third one singled out correctly 31 of the same testing cases (93,9%). For the renal syndromes domain, the performance of HYCONES II - SMART was superior to the one presented by the CNM version (p < 0,05), and equal to the performance presented by the Simplified Fuzzy ARTMAP version. SMART and Simplified Fuzzy ARTMAP singled out correctly 108 of the 127 testing cases (85%), while the CNM version pointed out correctly 95 of the same 127 testing cases (74.8%). Finally, it was observed that the neural networks generated by HYCONES II - SMART had a similar content to the networks generated by CNM and to the knowledge graphs elicited from multiple experts. The main contributions of this dissertation are: the design, implementation and validation of the Simplified Fuzzy ARTMAP and SMART models. The latter one, however, stands out for its learning mechanism, which provides a higher semantic value to the recognition categories, when compared to the categories formed by conventional ART models. This important enhancement is obtained by incorporating specificity and relevance concepts to ART's dynamics. This dissertation, however, represents not only the design and validation of two new connectionist models, but also, the enrichment of HYCONES. This is obtained through the continuation of a previous MSc dissertation, under the same supervision supervision. From the present work, therefore, it is given to the knowledge engineering, the choice among three different neural networks: CNM, Semantic ART and Simplified Fuzzy ARTMAP, all of which, display good performance. Indeed, the first and second models, in contrast to the third, support the context in a semantic way.