ONNIS-GI: uma rede neural oscilatória para segmentação de imagens implementada em arquitetura maciçamente paralela

A presente tese apresenta a concepção de uma rede neural oscilatória e sua realização em arquitetura maciçamente paralela, a qual é adequada à implementação de chips de visão digitais para segmentação de imagens. A rede proposta, em sua versão final, foi denominada ONNIS-GI (Oscillatory Neural Network for Image Segmentation with Global Inhibition) e foi inspirada em uma rede denominada LEGION (Locally Excitatory Globally Inhibitory Oscillator Network), também de concepção recente. Inicialmente, é apresentada uma introdução aos procedimentos de segmentação de imagens, cujo objetivo é o de situar e enfatizar a importância do tema abordado dentro de um contexto abrangente, o qual inclui aplicações de visão artificial em geral. Outro aspecto abordado diz respeito à utilização de redes neurais artificiais em segmentação de imagens, enfatizando as denominadas redes neurais oscilatórias, as quais têm apresentado resultados estimulantes nesta área. A implementação de chips de visão, integrando sensores de imagens e redes maciçamente paralelas de processadores, é também abordada no texto, ressaltando o objetivo prático da nova rede neural proposta. No estudo da rede LEGION, são apresentados resultados de aplicações originais desenvolvidas em segmentação de imagens, nos quais é verificada sua propriedade de separação temporal dos segmentos. A versão contínua da rede, um arranjo paralelo de neurônios baseados em equações diferenciais, apresenta elevada complexidade computacional para implementação em hardware digital e muitos parâmetros, com procedimento de ajuste pouco prático. Por outro lado, sua arquitetura maciçamente paralela apresenta-se particularmente adequada à implementação de chips de visão analógicos com capacidade de segmentação de imagens. Com base nos bons resultados obtidos nas aplicações desenvolvidas, é proposta uma nova rede neural, em duas versões, ONNIS e ONNIS-GI, as quais suplantam a rede LEGION em diversos aspectos relativos à implementação prática. A estrutura dos elementos de processamento das duas versões da rede, sua implementação em arquitetura maciçamente paralela e resultados de simulações e implementações em FPGA são apresentados, demonstrando a viabilidade da proposta. Como resultado final, conclui-se que a rede ONNIS-GI apresenta maior apelo de ordem prática, sendo uma abordagem inovadora e promissora na solução de problemas de segmentação de imagens, possuindo capacidade para separar temporalmente os segmentos encontrados e facilitando a posterior identificação dos mesmos. Sob o ponto de vista prático, a nova rede pode ser utilizada para implementar chips de visão digitais com arquitetura maciçamente paralela, explorando a velocidade de tais topologias e apresentando também flexibilidade para implementação de procedimentos de segmentação de imagens mais sofisticados.

A Learning Function for Parameter Reduction in Spiking Neural Networks with Radial Basis Function

Spiking neural networks - networks that encode information in the timing of spikes - are arising as a new approach in the artificial neural networks paradigm, emergent from cognitive science. One of these new models is the pulsed neural network with radial basis function, a network able to store information in the axonal propagation delay of neurons. Learning algorithms have been proposed to this model looking for mapping input pulses into output pulses. Recently, a new method was proposed to encode constant data into a temporal sequence of spikes, stimulating deeper studies in order to establish abilities and frontiers of this new approach. However, a well known problem of this kind of network is the high number of free parameters - more that 15 - to be properly configured or tuned in order to allow network convergence. This work presents for the first time a new learning function for this network training that allow the automatic configuration of one of the key network parameters: the synaptic weight decreasing factor.

Técnicas de inteligência artificial para a geração dinâmica de set points para uma coluna de destilação

Artificial Intelligence techniques are applied to improve performance of a simulated oil distillation system. The chosen system was a debutanizer column. At this process, the feed, which comes to the column, is segmented by heating. The lightest components become steams, by forming the LPG (Liquefied Petroleum Gas). The others components, C5+, continue liquid. In the composition of the LPG, ideally, we have only propane and butanes, but, in practice, there are contaminants, for example, pentanes. The objective of this work is to control pentane amount in LPG, by means of intelligent set points (SP s) determination for PID controllers that are present in original instrumentation (regulatory control) of the column. A fuzzy system will be responsible for adjusting the SP's, driven by the comparison between the molar fraction of the pentane present in the output of the plant (LPG) and the desired amount. However, the molar fraction of pentane is difficult to measure on-line, due to constraints such as: long intervals of measurement, high reliability and low cost. Therefore, an inference system was used, based on a multilayer neural network, to infer the pentane molar fraction through secondary variables of the column. Finally, the results shown that the proposed control system were able to control the value of pentane molar fraction under different operational situations

Análise de um controlador baseado no Jacobiano estimado da planta através de uma rede neural

This work presents an analysis of the control law based on an indirect hybrid scheme using neural network, initially proposed for O. Adetona, S. Sathanathan and L. H. Keel. Implementations of this control law, for a level plant of second order, was resulted an oscillatory behavior, even if the neural identifier has converged. Such results had motivated the investigation of the applicability of that law. Starting from that, had been made stability mathematical analysis and several implementations, with simulated plants and with real plants, for analyze the problem. The analysis has been showed the law was designed being despised some components of dynamic of the plant to be controlled. Thus, for plants that these components have a significant influence in its dynamic, the law tends to fail

Comparative study between RBF and radial-PPS neural networks

The study of function approximation is motivated by the human limitation and inability to register and manipulate with exact precision the behavior variations of the physical nature of a phenomenon. These variations are referred to as signals or signal functions. Many real world problem can be formulated as function approximation problems and from the viewpoint of artificial neural networks these can be seen as the problem of searching for a mapping that establishes a relationship from an input space to an output space through a process of network learning. Several paradigms of artificial neural networks (ANN) exist. Here we will be investigated a comparative of the ANN study of RBF with radial Polynomial Power of Sigmoids (PPS) in function approximation problems. Radial PPS are functions generated by linear combination of powers of sigmoids functions. The main objective of this paper is to show the advantages of the use of the radial PPS functions in relationship traditional RBF, through adaptive training and ridge regression techniques.

Application of neural networks to identify features of dynamical grounding systems

The accurate identification of features of dynamical grounding systems are extremely important to define the operational safety and proper functioning of electric power systems. Several experimental tests and theoretical investigations have been carried out to obtain characteristics and parameters associated with the technique of grounding. The grounding system involves a lot of non-linear parameters. This paper describes a novel approach for mapping characteristics of dynamical grounding systems using artificial neural networks. The network acts as identifier of structural features of the grounding processes. So that output parameters can be estimated and generalized from an input parameter set. The results obtained by the network are compared with other approaches also used to model grounding systems.

A neural system to robust Nonlinear optimization subject to disjoint and constrained sets

The ability of neural networks to realize some complex nonlinear function makes them attractive for system identification. This paper describes a novel method using artificial neural networks to solve robust parameter estimation problems for nonlinear models with unknown-but-bounded errors and uncertainties. More specifically, a modified Hopfield network is developed and its internal parameters are computed using the valid-subspace technique. These parameters guarantee the network convergence to the equilibrium points. A solution for the robust estimation problem with unknown-but-bounded error corresponds to an equilibrium point of the network. Simulation results are presented as an illustration of the proposed approach.

An efficient neural approach to economic load dispatch in power systems

A neural approach to solve the problem defined by the economic load dispatch in power systems is presented in this paper, Systems based on artificial neural networks have high computational rates due to the use of a massive number of simple processing elements and the high degree of connectivity between these elements the ability of neural networks to realize some complex nonlinear function makes them attractive for system optimization the neural networks applyed in economic load dispatch reported in literature sometimes fail to converge towards feasible equilibrium points the internal parameters of the modified Hopfield network developed here are computed using the valid-subspace technique These parameters guarantee the network convergence to feasible quilibrium points, A solution for the economic load dispatch problem corresponds to an equilibrium point of the network. Simulation results and comparative analysis in relation to other neural approaches are presented to illustrate efficiency of the proposed approach.

Uma análise da aplicação do modelo de Rede Neural RePART em Comitês de classificadores

RePART (Reward/Punishment ART) is a neural model that constitutes a variation of the Fuzzy Artmap model. This network was proposed in order to minimize the inherent problems in the Artmap-based model, such as the proliferation of categories and misclassification. RePART makes use of additional mechanisms, such as an instance counting parameter, a reward/punishment process and a variable vigilance parameter. The instance counting parameter, for instance, aims to minimize the misclassification problem, which is a consequence of the sensitivity to the noises, frequently presents in Artmap-based models. On the other hand, the use of the variable vigilance parameter tries to smoouth out the category proliferation problem, which is inherent of Artmap-based models, decreasing the complexity of the net. RePART was originally proposed in order to minimize the aforementioned problems and it was shown to have better performance (higer accuracy and lower complexity) than Artmap-based models. This work proposes an investigation of the performance of the RePART model in classifier ensembles. Different sizes, learning strategies and structures will be used in this investigation. As a result of this investigation, it is aimed to define the main advantages and drawbacks of this model, when used as a component in classifier ensembles. This can provide a broader foundation for the use of RePART in other pattern recognition applications

Classificador neural híbrido para imagens obtidas por sensoriamento remoto

Remote sensing is one technology of extreme importance, allowing capture of data from the Earth's surface that are used with various purposes, including, environmental monitoring, tracking usage of natural resources, geological prospecting and monitoring of disasters. One of the main applications of remote sensing is the generation of thematic maps and subsequent survey of areas from images generated by orbital or sub-orbital sensors. Pattern classification methods are used in the implementation of computational routines to automate this activity. Artificial neural networks present themselves as viable alternatives to traditional statistical classifiers, mainly for applications whose data show high dimensionality as those from hyperspectral sensors. This work main goal is to develop a classiffier based on neural networks radial basis function and Growing Neural Gas, which presents some advantages over using individual neural networks. The main idea is to use Growing Neural Gas's incremental characteristics to determine the radial basis function network's quantity and choice of centers in order to obtain a highly effective classiffier. To demonstrate the performance of the classiffier three studies case are presented along with the results.

Artificial immune systems for classification of petroleum well drilling operations

This paper presents two approaches of Artificial Immune System for Pattern Recognition (CLONALG and Parallel AIRS2) to classify automatically the well drilling operation stages. The classification is carried out through the analysis of some mud-logging parameters. In order to validate the performance of AIS techniques, the results were compared with others classification methods: neural network, support vector machine and lazy learning.

Designing a modified Hopfield network to solve an Economic Dispatch problem with nonlinear cost function

Economic Dispatch (ED) problems have recently been solved by artificial neural networks approaches. In most of these dispatch models, the cost function must be linear or quadratic. Therefore, functions that have several minimum points represent a problem to the simulation since these approaches have not accepted nonlinear cost function. Another drawback pointed out in the literature is that some of these neural approaches fail to converge efficiently towards feasible equilibrium points. This paper discusses the application of a modified Hopfield architecture for solving ED problems defined by nonlinear cost function. The internal parameters of the neural network adopted here are computed using the valid-subspace technique, which guarantees convergence to equilibrium points that represent a solution for the ED problem. Simulation results and a comparative analysis involving a 3-bus test system are presented to illustrate efficiency of the proposed approach.

Neural networks in chemosystematic studies of asteraceae: A classification based on a dichotomic approach

This paper describes the application of artificial neural nets as an alternative and efficient method for the classification of botanical taxa based on chemical data (chemosystematics). A total of 28,000 botanical occurrences of chemical compounds isolated from the Asteraceae family were chosen from the literature, and grouped by chemical class for each species. Four tests were carried out to differentiate and classify different botanical taxa. The qualifying capacity of the artificial neural nets was dichotomically tested at different hierarchical levels of the family, such as subfamilies and groups of Heliantheae subtribes. Furthermore, two specific subtribes of the Heliantheae and two genera of one of these subtribes were also tested. In general, the artificial neural net gave rise to good results, with multiple-correlation values R > 0.90. Hence, it was possible to differentiate the dichotomic character of the botanical taxa studied.

Harmonic identification using parallel neural networks in single-phase systems

Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq)