Na visão computacional, a detecção de objetos é uma tarefa que tem signifgificativa importância. Podemos verificar isto através da existência de inúmeros métodos propostos na literatura. Cada um destes métodos se apóia em algumas características presentes na imagem para alcançar um desempenho eficiente. Considerando ambientes que utilizam cores para determinação de objetos presentes em uma imagem, é possível utilizá-las como características que permitam detectar os objetos. Neste trabalho, são investigados dois classificadores de cores. O primeiro é baseado em limiarização no espaço HSV e o segundo é constituído de um mapa auto-organizável para classificação dos pixels no espaço RGB. Objetivando a construção de um sistema classificador de cores eficiiente, capaz de processar vídeo em tempo real, é proposta uma técnica que se baseia no conceito de quantização. Outro aspecto investigado foi a detecção de movimento para evitar o processamento de pontos indesejados. O desempenho do sistema de classificação de cores é avaliado em um ambiente de futebol de robôs da categoria Mirosot, que é um ambiente dinâmico e que exige que todo o processamento da imagem seja rápido de modo a detectar corretamente todos objetos presentes em cada quadro. Os resultados mostram que o classificador de cores é capaz de detectar todos objetos no ambiente de futebol de robôs, sendo cada quadro processado em menos de 30 milisegundos, tornando o sistema desenvolvido muito adequado ao processamento de vídeo / In computer vision, the detection of objects is a task of great importance. We can verify this by the existence of several methods proposed in the literature. Each one of these methods is based on some characteristics present in the image to reach an eficient performance. Considering environments that make use of colors for determining the objects present in a image, it is possible to utilize them as the characteristics that allow to detect the objects. In this work, two color classifiers are investigated. The first one is based on the thresholding in the HSV space and the second is constituted by a self-organizing map for classifying of pixels in the RGB space. Aiming to construct an eficient color classifier able to process video in real time, it is proposed a technique that is based on the quantization concept. It is also investigated the detection of movement to avoid processing undesired points. The performance of the color classifier system is validated in a MIROSOT robot soccer environment, which is a dynamic environment, requiring that all image processing be very fast in order to detect all the objects present in each frame. The results show that the color classifier system is able to detect correctly all objects present in the robot soccer environment, processing each frame in less than 30 milliseconds, turning the developed system very appropriate for real time video processing

O aprendizado de máquina consiste em conceitos e técnicas que permitem aos computadores melhorar seu desempenho com a experiência, ou em outras palavras, aprender com dados. Duas de suas principais categorias são o aprendizado não-supervisionado e o semissupervisionado, que respectivamente consistem em inferir padrões em bases cujos dados não têm rótulo (classe) e classificar dados em bases parcialmente rotuladas. Embora muito estudado, trata-se de um campo repleto de desafios e com muitos tópicos abertos. Sistemas dinâmicos coletivos, por sua vez, são sistemas constituídos por muitos indivíduos, cada qual um sistema dinâmico por si só, de modo que todos eles agem coletivamente, ou seja, a ação de cada indivíduo é influenciada pela ação dos vizinhos. Uma característica notável desses sistemas é que padrões globais podem surgir espontaneamente das interações locais entre os indivíduos, fenômeno conhecido como emergência. Os desafios intrínsecos e a relevância do tema vêm motivando sua pesquisa em diversos ramos da ciência e da engenharia. Este trabalho de doutorado consiste no desenvolvimento e análise de modelos dinâmicos coletivos para o aprendizado de máquina, especificamente suas categorias não-supervisionada e semissupervisionada. As tarefas de segmentação de imagens e de detecção de comunidades em redes, que de certo modo podem ser entendidas como tarefas do aprendizado de máquina, são também abordadas. Em especial, desenvolvem-se modelos nos quais a movimentação dos objetos é determinada pela localização e velocidade de seus vizinhos. O sistema dinâmico assim modelado é então conduzido a um estado cujo padrão formado por seus indivíduos realça padrões subjacentes do conjunto de dados. Devido ao seu caráter auto-organizável, os modelos aqui desenvolvidos são robustos e as informações geradas durante o processo (valores das variáveis do sistema) são ricas e podem, por exemplo, revelar características para realizar soft labeling e determinar classes sobrepostas. / Machine learning consists of concepts and techniques that enable computers to improve their performance with experience, i.e., learn from data. Unsupervised and semi-supervised learning are important categories of machine learning, which respectively consists of inferring patterns in datasets whose data have no label (class) and classifying data in partially-labeled datasets. Although intensively studied, machine learning is still a field full of challenges and with many open topics. Collective dynamical systems, in turn, are systems made of a large group of individuals, each one a dynamical system by itself, such that all of them behave collectively, i.e., the action of each individual is influenced by the action of its neighbors. A remarkable feature of those systems is that global patterns may spontaneously emerge from the local interactions among individuals, a phenomenon known as emergence. Their relevance and intrinsic challenges motivate research in various branches of science and engineering. In this doctorate research, we develop and analyze collective dynamical models for their usage in machine-learning tasks, specifically unsupervised and semi-supervised ones. Image segmentation and network community detection are also addressed, as they are related to machine learning as well. In particular, we propose to work on models in which the objects motion is determined by the location and velocity of their neighbors. By doing so, the dynamical system reaches a configuration in which the patterns developed by the set of individuals highlight underlying patterns of the dataset. Due to their self-organizing nature, it is also expected that the models can be robust and the information generated during the process (values of the system variables) can be rich and reveal, for example, features to perform soft labeling and determine overlapping classes.

Indiana University-Purdue University Indianapolis (IUPUI) / In recent years, Convolutional Neural Networks (CNNs) have become the state-of- the-art method for object detection and classi cation in the eld of machine learning and arti cial intelligence. In contrast to a fully connected network, each neuron of a convolutional layer of a CNN is connected to fewer selected neurons from the previous layers and kernels of a CNN share same weights and biases across the same input layer dimension. These features allow CNN architectures to have fewer parameters which in turn reduces calculation complexity and allows the network to be implemented in low power hardware. The accuracy of a CNN depends mostly on the number of images used to train the network, which requires a hundred thousand to a million images. Therefore, a reduced training alternative called transfer learning is used, which takes advantage of features from a pre-trained network and applies these features to the new problem of interest. This research has successfully developed a new CNN based on the pre-trained CIFAR-10 network and has used transfer learning on a new problem to classify road edges. Two network sizes were tested: 32 and 16 Neuron inputs with 239 labeled Google street view images on a single CPU. The result of the training gives 52.8% and 35.2% accuracy respectively for 250 test images. In the second part of the research, High Level Synthesis (HLS) hardware model of the network with 16 Neuron inputs is created for the Zynq 7000 FPGA. The resulting circuit has 34% average FPGA utilization and 2.47 Watt power consumption. Recommendations to improve the classi cation accuracy with deeper network and ways to t the improved network on the FPGA are also mentioned at the end of the work.

京都大学 / 0048 / 新制・課程博士 / 博士(医学) / 甲第19616号 / 医博第4123号 / 新制||医||1015(附属図書館) / 32652 / 京都大学大学院医学研究科医学専攻 / (主査)教授 佐藤 俊哉, 教授 富樫 かおり, 教授 藤渕 航 / 学位規則第4条第1項該当 / Doctor of Medical Science / Kyoto University / DFAM

Proteins are among the most versatile organic macromolecules essential for living systems and present in almost all biological processes. Cancer is associated with mutations that either enhance or disrupt the conformation of proteins. These mutations have been shown to accumulate in specific regions of a proteins three dimensional structure. In this thesis, the aim is to find connections that secondary structure elements make and explore them using a self-organizing map (SOM). The detection of these connections is done by first mapping the three-dimensional structure onto a novice type of distance matrix that also incorporates chemical information, and then deploying a density-based clustering algorithm. The connections found are mapped onto the SOM and later analyzed in order to see if highly mutated connections are more common among certain SOM-nodes. This was tested with an ANOVA that indicated that there are indeed mutational asymmetries among the nodes. By further analyzing the map it could also be stated that certain nodes were to a large extent activated by connections from genes associated with cancer. / Proteiner tillhör några av de mest mångsidiga organiska makromolekylerna, och är direkt nödvändiga för alla levande system och biologiska processer. Cancer orsakas av mutationer som antingen förstärker eller stör strukturen hos proteinet. Dessa mutationer tenderar att att samlas i specifika områden av proteinets tredimensionella struktur. I den här rapporten är målet att hitta kopplingar som sekundärstrukturselement skapar, och utforska dem med hjälp av en självorganiserande karta. Dessa kopplingar finnes genom att först skapa en tvådimensionell representation av proteinets tredimensionella struktur, och sedan använda en densitetsbaserad klustringsalgoritm. De funna kopplingarna mappas till de olika neuronerna i kartan och analyseras sedan för att se om kopplingar med hög mutationsnivå är mer vanliga hos vissa neuron. För att undersöka detta användes ett ANOVA-test som visade att så var fallet. Genom att ytterligare studera kartan upptäcktes fynd som indikerade att vissa neuron i högre utsträckning var aktiverade av kopplingar som härstammar från gener vi vet är associerade med cancer.

Studies of animal spatial distributions typically use prior knowledge of animal habitat requirements and behavioral ecology to deduce the most likely explanations of observed habitat use. Animal-borne accelerometers can be used to distinguish behaviors which allows us to incorporate in situ behavior into our understanding of spatial distributions. Past research has focused on using supervised machine-learning, which requires a priori specification of behavior to identify signals whereas unsupervised approaches allow the model to identify as many signal types as permitted by the data. The following framework couples direct observation to behavioral clusters identified from unsupervised machine learning on a large accelerometry dataset. A behavioral profile was constructed to describe the proportion of behaviors observed per cluster and the framework was applied to an acceleration dataset collected from wild pigs (Sus scrofa). Although, most clusters represented combinations of behaviors, a leave-p-out validation procedure indicated this classification system accurately predicted new data.

No description available.

No description available.

<p>A cognitive radio network is a multi-user system, in which different radio units compete for limited resources in an opportunistic manner, interacting with each other for access to the available resources. The fact that both users and spectrum holes (i.e., under-utilized spectrum sub-bands) can come and go in a stochastic manner, makes a cognitive radio network a highly non- stationary, dynamic and challenging wireless environment. Finding robust decentralized resource-allocation algorithms, which are capable of achieving reasonably good solutions fast enough in order to guarantee an acceptable level of performance, is crucial in such an environment. In this thesis, a novel dynamic spectrum management (DSM) scheme for cognitive radio networks, termed the self-organizing dynamic spectrum management (SO-DSM), is described and its practical validity is demonstrated using computer simulations. In this scheme, CRs try to exploit the primary networks’ unused bands and establish link with neighbouring CRs using those bands. Inspired by human brain, the CRs extract and memorize primary network’s and other CRs’ activity patterns and create temporal channel assignments on sub-bands with no recent primary user activities using self-organizing maps (SOM) technique. The proposed scheme is decentralized and employs a simple learning rule with low complexity and minimal memory requirements. A software testbed was developed to simulate and study the proposed scheme. This testbed is capable of simulating CR network alongside of a cellular legacy network. In addition to SO-DSM, two other DSM schemes, namely centralized DSM and no-learning decentralized DSM, can be used for CR networks in this software testbed. The software testbed was deployed on parallel high capacity computing clusters from Sharcnet to perform large scale simulations of CR network. The simulation results show, comparing to centralized DSM and minority game DSM (MG-DSM), the SO-DSM decreases the probability of collision with primary users and also probability of CR link interruption significantly with a moderate decrease in CR network spectrum utilization.</p> / Doctor of Philosophy (PhD)

Publish/subscribe (P/S) is a communication paradigm of growing popularity for information dissemination in large-scale distributed systems. The weak coupling between information producers and consumers in P/S systems is attractive for loosely coupled and dynamic network infrastructures such as ad hoc networks. However, achieving end-to-end timeliness and reliability properties when P/S events are causally dependent is an open problem in ad hoc networks. In this thesis, we present, evaluate benefits of, and compare with past work, an architecture design that can effectively support timely and reliable delivery of events and causally related events in ad hoc environments, and especially in mobile ad hoc networks (MANETs). With observations from both realistic application model and simulation experiments, we reveal causal dependencies among events and their significance in a typical use notional system. We also examine and propose engineering methodologies to further tailor an event-based system to facilitate its self-reorganizing capability and self-reconfiguration. Our design features a two-layer structure, including novel distributed algorithms and mechanisms for P/S tree construction and maintenance. The trace-based experimental simulation studies illustrate our design's effectiveness in both cases with and without causal dependencies. / Master of Science