Os seres vivos têm uma impressionante capacidade de lidar com ambientes complexos com grandes quantidades de informações de forma muito autônoma. Isto os torna um modelo ideal para o desenvolvimento de sistemas artificiais bioinspirados. A rede neural artificial auto-organizada de Kohonen é um excelente exemplo de um sistema baseado nos modelos biológicos. Esta tese discutirá ilustrativamente o reconhecimento e a generalização de padrões em alta dimensão nos sistemas biológicos e como eles lidam com redução de dimensionalidade para otimizar o armazenamento e o acesso às informações memorizadas para fins de reconhecimento e categorização de padrões, mas apenas para contextualizar o tema com as propostas desta tese. As novas propostas desenvolvidas nesta tese são úteis para aplicações de extração não supervisionada de conhecimento a partir dos mapas auto-organizados. Trabalha-se sobre o modelo da Rede Neural de Kohonen, mas algumas das metodologias propostas também são aplicáveis com outras abordagens de redes neurais auto-organizadas. Será apresentada uma técnica de reconstrução visual dos neurônios do Mapa de Kohonen gerado pelo método híbrido PCA+SOM. Essa técnica é útil quando se trabalha com banco de dados de imagens. Propõe-se também um método para melhorar a representação dos dados do mapa SOM e discute-se o resultado do mapeamento SOM como uma generalização das informações do espaço de dados. Finalmente, apresenta-se um método de exploração de espaço de dados em alta dimensão de maneira auto-organizada, baseado no manifold dos dados, cuja proposta foi denominada Self Organizing Manifold Mapping (SOMM). São apresentados os resultados computacionais de ensaios realizados com cada uma das propostas acima e eles são avaliados as com métricas de qualidade conhecidas, além de uma nova métrica que está sendo proposta neste trabalho. / Living beings have an amazing capacity to deal with complex environments with large amounts of information autonomously. They are the perfect model for bioinspired artificial system development. The artificial neural network developed by Kohonen is an excellent example of a system based on biological models. In this thesis, we will discuss illustratively pattern recognition and pattern generalization in high dimensional data space by biological system. Then, a brief discussion of how they manage dimensionality reduction to optimize memory space and speed up information access in order to categorize and recognize patterns. The new proposals developed in this thesis are useful for applications of unsupervised knowledge extraction using self-organizing maps. The proposals use Kohonens model. However, any self-organizing neural network in general can also use the proposed techniques. It will be presented a visual reconstruction technique for Kohonens neurons, which was generated by hybrid method PCA+SOM. This technique is useful when working with images database. It is also proposed a method for improving the representation of SOMs map and discussing the result of the SOMs mapping as a generalization of the information data space. Finally, it is proposed a method for exploring high dimension data space in a self-organized way on the data manifold. This new proposal was called Self Organizing Manifold Mapping (SOMM). We present the results of computational experiments on each of the above proposals and evaluate the results using known quality metrics, as well as a new metric that is being proposed in this thesis.

O avanço tecnológico atual está permitindo que as pessoas recebam cada vez mais informações visuais dos mais diferentes tipos, nas mais variadas mídias. Esse aumento fantástico está obrigando os pesquisadores e as indústrias a imaginar soluções para o armazenamento e recuperação deste tipo de informação, pois nossos computadores ainda utilizam, apesar dos grandes avanços nessa área, um sistema de arquivos imaginado há décadas, quando era natural trabalhar com informações meramente textuais. Agora, nos deparamos com novos problemas: Como encontrar uma paisagem específica em um banco de imagens, em que trecho de um filme aparece um cavalo sobre uma colina, em que parte da fotografia existe um gato, como fazer um robô localizar um objeto em uma cena, entre outras necessidades. O objetivo desse trabalho é propor uma arquitetura de rede neural artificial que permita o reconhecimento de objetos genéricos e de categorias em banco de imagens digitais, de forma que se possa recuperar imagens específicas a partir da descrição da cena fornecida pelo usuário. Para que esse objetivo fosse alcançado, foram utilizadas técnicas de Visão Computacional e Processamento de Imagens na etapa de extração de feições de baixo nível e de Redes Neurais(Mapas Auto-Organizáveis de Kohonen) na etapa de agrupamento de classes de objetos. O resultado final desse trabalho pretende ser um embrião para um sistema de reconhecimento de objetos mais genérico, que possa ser estendido para a criação de indices de forma automática ou semi-automática em grandes bancos de imagens. / The current technological progress allows people to receive more and more visual information of the most different types, in different medias. This huge augmentation of image availability forces researchers and industries to propose efficient solutions for image storage and recovery. Despite the extraordinary advances in computational power, the data files system remain the same for decades, when it was natural to deal only with textual information. Nowadays, new problems are in front of us in this field. For instance, how can we find an specific landscape in a image database, in which place of a movie there is a horse on a hill, in which part of a photographic picture there is a cat, how can a robot find an object in a scene, among other queries. The objective of this work is to propose an Artificial Neural Network (ANN) architecture that performs the recognition of generic objects and object’s categories in a digital image database. With this implementation, it becomes possible to do image retrieval through the user´s scene description. To achieve our goal, we have used Computer Vision and Image Processing techniques in low level features extraction and Neural Networks (namely Kohonen’s Self-Organizing Maps) in the phase of object classes clustering. The main result of this work aims to be a seed for a more generic object recognition system, which can be extended to the automatic or semi-automatic index creation in huge image databases.

Chan Chi Hang. / Thesis submitted in: August 2004. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2005. / Includes bibliographical references (leaves 105-114). / Abstracts in English and Chinese. / Abstract --- p.i / Acknowledgement --- p.iv / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- Problem Statement --- p.3 / Chapter 1.2 --- Major Contributions --- p.5 / Chapter 1.3 --- Publication List --- p.6 / Chapter 1.4 --- Thesis Organization --- p.7 / Chapter 2 --- Background Survey --- p.9 / Chapter 2.1 --- Relevance Feedback Framework --- p.9 / Chapter 2.1.1 --- Relevance Feedback Types --- p.11 / Chapter 2.1.2 --- Data Distribution --- p.12 / Chapter 2.1.3 --- Training Set Size --- p.14 / Chapter 2.1.4 --- Inter-Query Learning and Intra-Query Learning --- p.15 / Chapter 2.2 --- History of Relevance Feedback Techniques --- p.16 / Chapter 2.3 --- Relevance Feedback Approaches --- p.19 / Chapter 2.3.1 --- Vector Space Model --- p.19 / Chapter 2.3.2 --- Ad-hoc Re-weighting --- p.26 / Chapter 2.3.3 --- Distance Optimization Approach --- p.29 / Chapter 2.3.4 --- Probabilistic Model --- p.33 / Chapter 2.3.5 --- Bayesian Approach --- p.39 / Chapter 2.3.6 --- Density Estimation Approach --- p.42 / Chapter 2.3.7 --- Support Vector Machine --- p.48 / Chapter 2.4 --- Presentation Set Selection --- p.52 / Chapter 2.4.1 --- Most-probable strategy --- p.52 / Chapter 2.4.2 --- Most-informative strategy --- p.52 / Chapter 3 --- Biased Support Vector Machine for Content-Based Image Retrieval --- p.57 / Chapter 3.1 --- Motivation --- p.57 / Chapter 3.2 --- Background --- p.58 / Chapter 3.2.1 --- Regular Support Vector Machine --- p.59 / Chapter 3.2.2 --- One-class Support Vector Machine --- p.61 / Chapter 3.3 --- Biased Support Vector Machine --- p.63 / Chapter 3.4 --- Interpretation of parameters in BSVM --- p.67 / Chapter 3.5 --- Soft Label Biased Support Vector Machine --- p.69 / Chapter 3.6 --- Interpretation of parameters in Soft Label BSVM --- p.73 / Chapter 3.7 --- Relevance Feedback Using Biased Support Vector Machine --- p.74 / Chapter 3.7.1 --- Advantages of BSVM in Relevance Feedback . . --- p.74 / Chapter 3.7.2 --- Relevance Feedback Algorithm By BSVM --- p.75 / Chapter 3.8 --- Experiments --- p.78 / Chapter 3.8.1 --- Synthetic Dataset --- p.80 / Chapter 3.8.2 --- Real-World Dataset --- p.81 / Chapter 3.8.3 --- Experimental Results --- p.83 / Chapter 3.9 --- Conclusion --- p.86 / Chapter 4 --- Self-Organizing Map-based Inter-Query Learning --- p.88 / Chapter 4.1 --- Motivation --- p.88 / Chapter 4.2 --- Algorithm --- p.89 / Chapter 4.2.1 --- Initialization and Replication of SOM --- p.89 / Chapter 4.2.2 --- SOM Training for Inter-Query Learning --- p.90 / Chapter 4.2.3 --- Incorporate with Intra-Query Learning --- p.92 / Chapter 4.3 --- Experiments --- p.93 / Chapter 4.3.1 --- Synthetic Dataset --- p.95 / Chapter 4.3.2 --- Real-World Dataset --- p.95 / Chapter 4.3.3 --- Experimental Results --- p.97 / Chapter 4.4 --- Conclusion --- p.98 / Chapter 5 --- Conclusion --- p.102 / Bibliography --- p.104

The enormous increase in the production of electronic data in today's information era has led to more challenges in analysing and understanding of the data. The rise in the innovations of technology devices, computers and the Internet has made it much easier to collect and store different kind of data ranging from personal, medical, financial, and scientific data. The growth in the amount of the generated data has introduced the term “Big Data” to describe this extremely high-dimensional and yet complex data. Making sense of the generated data sets is of great importance for the discovery of meaningful information that can be used to support decision making. Data mining techniques have been designed as a process for ex-ploring these data sets to extract meaning for decision making. An essential phase of the data mining procedure is the data transformation that involves the selection of input parameters. Selecting the right input parameters has a great impact on the performance of machine learning algorithms. Currently, there are existing manual statistical methods that are used for this task, but these are difficult to use, time consuming and require an expert. Automated data analysis is the initial step to relieve this burden from humans, through the provision of a systematic procedure of inspecting, transforming and modelling data for knowledge discovery. This project presents a novel method that exploits the power of self-organization for a sys-tematic procedure of conducting and inspecting data classification, with the identification of input parameters that are important for the process. The developed method can be used on different classification problems with practical application in various areas such as health con-dition monitoring in health care, machinery fault detection and analysis, and financial instrument analysis among others.

Yip King-Fung. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2003. / Includes bibliographical references (leaves 84-89). / Abstracts in English and Chinese. / Abstract --- p.i / 論文摘要 --- p.iii / Table of Contents --- p.iv / List of Abbreviations --- p.vi / Acknowledgements --- p.vii / Chapter 1. --- Introduction --- p.1 / Chapter 1.1. --- Background --- p.1 / Chapter 1.2. --- Motivation --- p.3 / Chapter 1.3. --- Thesis Organization --- p.4 / Chapter 2. --- A Survey of Content-based Image Retrieval --- p.5 / Chapter 2.1. --- Text-based Image Retrieval --- p.5 / Chapter 2.2. --- Content-Based Image Retrieval --- p.7 / Chapter 2.2.1. --- Content-Based Image Retrieval Systems --- p.7 / Chapter 2.2.2. --- Query Methods --- p.9 / Chapter 2.2.3. --- Image Features --- p.11 / Chapter 2.2.4. --- Summary --- p.16 / Chapter 3. --- Visual Thesaurus using SOM --- p.17 / Chapter 3.1. --- Algorithm --- p.17 / Chapter 3.1.1. --- Image Representation --- p.17 / Chapter 3.1.2. --- Self-Organizing Map --- p.21 / Chapter 3.2. --- Preliminary Experiment --- p.27 / Chapter 3.2.1. --- Feature differences --- p.27 / Chapter 3.2.2. --- Labeling differences --- p.30 / Chapter 4. --- Experiment --- p.33 / Chapter 4.1. --- Subjects --- p.33 / Chapter 4.2. --- Apparatus --- p.33 / Chapter 4.2.1. --- Systems --- p.33 / Chapter 4.2.2. --- Test Databases --- p.33 / Chapter 4.3. --- Procedure --- p.34 / Chapter 4.3.1. --- Description --- p.35 / Chapter 4.3.2. --- SOM (text) --- p.36 / Chapter 4.3.3. --- SOM (image) --- p.38 / Chapter 4.3.4. --- QBE (text) --- p.40 / Chapter 4.3.5. --- QBE (image) --- p.42 / Chapter 4.3.6. --- Questionnaire --- p.44 / Chapter 4.3.7. --- Experiment Flow --- p.45 / Chapter 4.4. --- Results --- p.46 / Chapter 4.5. --- Discussion --- p.51 / Chapter 5. --- Quantizing Color Histogram --- p.55 / Chapter 5.1. --- Algorithm --- p.56 / Chapter 5.1.1. --- Codebook Generation Phrase --- p.57 / Chapter 5.1.2. --- Histogram Generation Phrase --- p.66 / Chapter 5.2. --- Experiment --- p.67 / Chapter 5.2.1. --- Test Database --- p.67 / Chapter 5.2.2. --- Evaluation Methods --- p.67 / Chapter 5.2.3. --- Results and Discussion --- p.69 / Chapter 5.2.4. --- Summary --- p.74 / Chapter 6. --- Relevance Feedback --- p.75 / Chapter 6.1. --- Relevance Feedback in Text Information Retrieval --- p.75 / Chapter 6.2. --- Relevance Feedback in Multimedia Information Retrieval --- p.76 / Chapter 6.3. --- Relevance Feedback in Visual Thesaurus --- p.76 / Chapter 7. --- Conclusions --- p.80 / Chapter 7.1. --- Applications --- p.81 / Chapter 7.2. --- Future Directions --- p.81 / Chapter 7.2.1. --- SOM Generation --- p.81 / Chapter 7.2.2. --- Hybrid Architecture --- p.82 / References --- p.84

Genetic algorithms are some of the most flexible among optimization methods. Because of their low requirements on input data, they are able to solve a wide array of problems. The flexibility is balanced by their lower effectiveness. When compared to more specialized methods, their results are inferior. This thesis examines the possibility of increasing their effectiveness by means of controlling their run by an artificial neural network. Presented inside are means of controlling a run of a genetic algorithm by a self-organizing map. The thesis contains an algorithm proposal, a prototype implementation of such algorithm and a series of tests to assess its efficiency. While the results on benchmark functions show some positive properties, the problems of greater complexity yield less optimistic results.

Os seres vivos têm uma impressionante capacidade de lidar com ambientes complexos com grandes quantidades de informações de forma muito autônoma. Isto os torna um modelo ideal para o desenvolvimento de sistemas artificiais bioinspirados. A rede neural artificial auto-organizada de Kohonen é um excelente exemplo de um sistema baseado nos modelos biológicos. Esta tese discutirá ilustrativamente o reconhecimento e a generalização de padrões em alta dimensão nos sistemas biológicos e como eles lidam com redução de dimensionalidade para otimizar o armazenamento e o acesso às informações memorizadas para fins de reconhecimento e categorização de padrões, mas apenas para contextualizar o tema com as propostas desta tese. As novas propostas desenvolvidas nesta tese são úteis para aplicações de extração não supervisionada de conhecimento a partir dos mapas auto-organizados. Trabalha-se sobre o modelo da Rede Neural de Kohonen, mas algumas das metodologias propostas também são aplicáveis com outras abordagens de redes neurais auto-organizadas. Será apresentada uma técnica de reconstrução visual dos neurônios do Mapa de Kohonen gerado pelo método híbrido PCA+SOM. Essa técnica é útil quando se trabalha com banco de dados de imagens. Propõe-se também um método para melhorar a representação dos dados do mapa SOM e discute-se o resultado do mapeamento SOM como uma generalização das informações do espaço de dados. Finalmente, apresenta-se um método de exploração de espaço de dados em alta dimensão de maneira auto-organizada, baseado no manifold dos dados, cuja proposta foi denominada Self Organizing Manifold Mapping (SOMM). São apresentados os resultados computacionais de ensaios realizados com cada uma das propostas acima e eles são avaliados as com métricas de qualidade conhecidas, além de uma nova métrica que está sendo proposta neste trabalho. / Living beings have an amazing capacity to deal with complex environments with large amounts of information autonomously. They are the perfect model for bioinspired artificial system development. The artificial neural network developed by Kohonen is an excellent example of a system based on biological models. In this thesis, we will discuss illustratively pattern recognition and pattern generalization in high dimensional data space by biological system. Then, a brief discussion of how they manage dimensionality reduction to optimize memory space and speed up information access in order to categorize and recognize patterns. The new proposals developed in this thesis are useful for applications of unsupervised knowledge extraction using self-organizing maps. The proposals use Kohonens model. However, any self-organizing neural network in general can also use the proposed techniques. It will be presented a visual reconstruction technique for Kohonens neurons, which was generated by hybrid method PCA+SOM. This technique is useful when working with images database. It is also proposed a method for improving the representation of SOMs map and discussing the result of the SOMs mapping as a generalization of the information data space. Finally, it is proposed a method for exploring high dimension data space in a self-organized way on the data manifold. This new proposal was called Self Organizing Manifold Mapping (SOMM). We present the results of computational experiments on each of the above proposals and evaluate the results using known quality metrics, as well as a new metric that is being proposed in this thesis.

O elevado custo de manutenção nos ambientes industriais motivou pesquisas de novas técnicas para melhorar as ações de reparos. Com a evolução tecnológica, principalmente da eletrônica, que proporcionou o uso de sistemas embarcados para melhorar as atividades de manutenção, estas agregaram inteligência e evoluíram para uma manutenção pró-ativa. Através de ferramentas de processamento de sinais, inteligência artificial e tolerância a falhas, surgiram novas abordagens para os sistemas de monitoramento a serviço da equipe de manutenção. Os ditos sistemas de manutenção inteligente, cuja tarefa é realizar testes em funcionamento (on-line) nos equipamentos industriais, promovem novos modelos de confiabilidade e disponibilidade. Tais sistemas são baseados nos conceitos de tolerância a falhas, e visam detectar, diagnosticar e predizer a ocorrência de falhas. Deste modo, fornece-se aos engenheiros de manutenção a informação antecipada do estado de comportamento do equipamento antes mesmo deste manifestar uma falha, reduzindo custos, aumentando a vida útil e tornando previsível o reparo. Para o desenvolvimento do sistema de manutenção inteligente objeto deste trabalho, foram estudadas técnicas de inteligência artificial (redes neurais artificiais), técnicas de projeto de sistemas embarcados e de prototipação em plataformas de hardware. No presente trabalho, a rede neural Mapas Auto-Organizáveis foi adotada como ferramenta base para detecção e diagnóstico de falhas. Esta foi prototipada numa plataforma de sistema embarcado baseada na tecnologia FPGA (Field Programmable Gate Array). Como estudo de caso, uma válvula elétrica utilizada em dutos para transporte de petróleo foi definida como aplicação alvo dos experimentos. Através de um modelo matemático, um conjunto de dados representativo do comportamento da válvula foi simulado e utilizado como entrada do sistema proposto. Estes dados visam o treinamento da rede neural e visam fornecer casos de teste para experimentação no sistema. Os experimentos executados em software validaram o uso da rede neural como técnica para detecção e diagnóstico de falhas em válvulas elétricas. Por fim, também realizou-se experimentos a fim de validar o projeto do sistema embarcado, comparando-se os resultado obtidos com este aos resultados obtidos a partir de testes em software. Os resultados revelam a escolha correta do uso da rede neural e o correto projeto do sistema embarcado para desempenhar as tarefas de detecção e diagnóstico de falhas em válvulas elétricas. / The high costs of maintenance in industrial environments have motivated research for new techniques to improve repair activities. The technological progress, especially in the electronics field, has provided for the use of embedded systems to improve repair, by adding intelligence to the system and turning the maintenance a proactive activity. Through tools like signal processing, artificial intelligence and fault-tolerance, new approaches to monitoring systems have emerged to serve the maintenance staff, leading to new models of reliability and availability. The main goal of these systems, also called intelligent maintenance systems, is to perform in-operation (on-line) test of industrial equipments. These systems are built based on fault-tolerance concepts, and used for the detection, the diagnosis and the prognosis of faults. They provide the maintenance engineers with information on the equipment behavior, prior to the occurrence of failures, reducing maintenance costs, increasing the system lifetime and making it possible to schedule repairing stops. To develop the intelligent maintenance system addressed in this dissertation, artificial intelligence (neural networks), embedded systems design and hardware prototyping techniques were studied. In this work, the neural network Self-Organizing Maps (SOM) was defined as the basic tool for the detection and the diagnosis of faults. The SOM was prototyped in an embedded system platform based on the FPGA technology (Field Programmable Gate Array). As a case study, the experiments were performed on an electric valve used in a pipe network for oil transportation. Through a mathematical model, a data set representative of the valve behavior was obtained and used as input to the proposed maintenance system. These data were used for neural network training and also provided test cases for system monitoring. The experiments were performed in software to validate the chosen neural network as the technique for the detection and diagnosis of faults in the electrical valve. Finally, experiments to validate the embedded system design were also performed, so as to compare the obtained results to those resulting from the software tests. The results show the correct choice of the neural network and the correct embedded systems design to perform the activities for the detection and diagnosis of faults in the electrical valve.

This work aims to improve the capability of accurate information extraction from high-dimensional data, with a specific neural learning paradigm, the Self-Organizing Map (SOM). The SOM is an unsupervised learning algorithm that can faithfully sense the manifold structure and support supervised learning of relevant information from the data. Yet open problems regarding SOM learning exist. We focus on the following two issues. (1) Evaluation of topology preservation. Topology preservation is essential for SOMs in faithful representation of manifold structure. However, in reality, topology violations are not unusual, especially when the data have complicated structure. Measures capable of accurately quantifying and informatively expressing topology violations are lacking. One contribution of this work is a new measure, the Weighted Differential Topographic Function ( WDTF ), which differentiates an existing measure, the Topographic Function ( TF ), and incorporates detailed data distribution as an importance weighting of violations to distinguish severe violations from insignificant ones. Another contribution is an interactive visual tool, TopoView, which facilitates the visual inspection of violations on the SOM lattice. We show the effectiveness of the combined use of the WDTF and TopoView through a simple two-dimensional data set and two hyperspectral images. (2) Learning multiple latent variables from high-dimensional data. We use an existing two-layer SOM-hybrid supervised architecture, which captures the manifold structure in its SOM hidden layer, and then, uses its output layer to perform the supervised learning of latent variables. In the customary way, the output layer only uses the strongest output of the SOM neurons. This severely limits the learning capability. We allow multiple, k , strongest responses of the SOM neurons for the supervised learning. Moreover, the fact that different latent variables can be best learned with different values of k motivates a new neural architecture, the Conjoined Twins, which extends the existing architecture with additional copies of the output layer, for preferential use of different values of k in the learning of different latent variables. We also automate the customization of k for different variables with the statistics derived from the SOM. The Conjoined Twins shows its effectiveness in the inference of two physical parameters from Near-Infrared spectra of planetary ices.

The perceptual reality of the music theoretical relation between the Neapolitan chord and the dominant / and the secondary dominant chord and its diatonic associate was investigated within the chord priming paradigm. In Experiment 1, expectation towards the dominant chord after the Neapolitan chord was observed in Turkish musicians and non-musicians with piano timbre. In Experiment 2, expectation towards the dominant chord after the Neapolitan chord was observed in European musicians but not in European non-musicians. In Experiment 3, Turkish non-musicians were tested with Shepard tones / but it was not possible to observe any priming effects. To understand effects of cultural background on the difference between the results of Experiments 1 and 2 further studies are necessary. In Experiments 4-5, the perceptual reality of the relation between the secondary dominant chord and its diatonic associate was investigated in Turkish non-musicians. In Experiment 4, chord sequences that included secondary dominant chords were played with Shepard tones / and they were scrambled with 2by2 scrambling algorithm. Experiment 5 was identical with Experiment 4, except chord sequences were played with the piano timbre. Experiment 6 was identical with Experiment 5, except chord sequences were scrambled with 4by4. However, in Experiments 4-6 detrimental effects of scrambling sequences that include secondary dominant chords on the priming of chords were not observed. Turkish non-musicians did perceive the relation between the secondary dominant chord and its diatonic associate. In neural network simulations of this thesis it was shown that statistical learning from the musical environment with self-organization could be achieved without committing the questionable assumptions of previous studies.