A teoria de redes neurais morfológicas e suas aplicações têm experimentado um crescimento contínuo e crescente nos últimos anos. Neste contexto, calcular o próximo estado de um neurônio, ou de uma camada, envolve uma das operações elementares da morfologia matemática. Nesta dissertação, forneceremos a caracterização de alguns modelos de redes neurais morfológicas, bem fundamentados pela teoria de morfologia matemática em reticulados completos, e também apresentaremos uma comparação do desempenho dos modelos em problemas de classificação; The theory of morphological neural networks and its applications have experiencied a steady and consistent growth in the last few years. In this setting, computing the next state of a neuron or performing the next layer computation involves one of the elementary operations of mathematical morphology. In this dissertation, we will provide a characterization of several morphological neural networks, well conduct by the theory of mathematical morphology over complete lattices, and we will also present a comparison of the performance of the models over classification problems

Redes neurais morfológicas(MNN) são redes neurais artificiais cujos nós executam operações elementares da morfologia matemática(MM). Vários modelos de MNNs e seus respectivos algoritmos de treinamentos têm sido propostos nos últimos anos, incluindo os perceptrons morfológicos(MPs), o perceptron morfológico com dendritos, as memórias associativas morfológicas (fuzzy), as redes neurais morfológicas modulares e as redes neurais de pesos compartilhados e regularizados. Aplicações de MNNs incluem reconhecimento de padrão, previsão de séries temporais, detecção de alvos, auto-localização e processamento de imagens hiperespectrais. Nesta tese, abordamos dois novos modelos de redes neurais morfológicas.O primeiro consiste em uma memória associativa fuzzy denominada KS-FAM, e o segundo representa uma nova versão do perceptron morfológico para problemas de classificação de múltiplas classes, denominado perceptron morfológico com aprendizagem competitiva(MP/CL). Para ambos modelos, investigamos e demonstramos várias propriedades. Em particular para a KS-FAM, caracterizamos as condições para que uma memória seja perfeitamente recordada, assim como a formada saída produzida ao apresentar um padrão de entrada qualquer. Provamos ainda que o algoritmo de treinamento do MP/CL converge em um número finito de passos e que a rede produzida independe da ordem com que os padrões de treinamento são apresentados. Além disso...

The Federal Aviation Administration (FAA) has been investigating early warning accident prevention systems in an effort to prevent runway collisions. One system in place is the Airport Movement Area Safety System (AMASS), developed under contract with the FAA. AMASS uses a linear prediction system to predict the position of an aircraft 5 to 30 seconds in the future. The system sounds an alarm to warn air traffic controllers if it foresees a potential accident. However, research done at MIT and Volpe National Transportation Systems Center has shown that neural networks more accurately predict the future position of aircraft. Neural networks are self-learning, and the time required for the optimization of safety logic will be minimized using neural networks. More accurate predictions of aircraft position will deliver earlier warnings to air traffic controllers while reducing the number of nuisance alerts. There are many factors to consider in designing an aircraft position prediction neural network, including history length, types of inputs and outputs, and applicable training data. This document chronicles the design, training, performance, and analysis of a position prediction neural network, and the presents the resulting optimal neural network for the AMASS System. Additionally...

This paper presents some intermediate results on fingerprint classification adopting a neural network as decision stage, in order to evaluate the performance of a discrete wavelet transform as feature extraction technique. Some issues on the image acquisition, preprocessing and segmentation are also discussed.; 5000

This research analyses the suitability of back-propagation artifical neural networks (ANNs) for modelling multivariate water quality time series. The ANNs are successfully applied to two case studies, the long-term forcasting of salinity and the modelling of blue-green algae, in the River Murray, Australia.; Thesis (Ph.D.)--University of Adelaide, Dept. of Civil and Environmental Engineering, 1996?; Corrigenda attached to back end paper.; Bibliography: p. 526-559.; xxx, 559 p. : ill. ; 30 cm.; Title page, contents and abstract only. The complete thesis in print form is available from the University Library.

A methodology is formulated for the successful design and implementation of artificial neural networks (ANN) models for water resources applications. Attention is paid to each of the steps that should be followed in order to develop an optimal ANN model; including when ANNs should be used in preference to more conventional statistical models; dividing the available data into subsets for modelling purposes; deciding on a suitable data transformation; determination of significant model inputs; choice of network type and architecture; selection of an appropriate performance measure; training (optimisation) of the networks weights; and, deployment of the optimised ANN model in an operational environment. The developed methodology is successfully applied to two water resorces case studies; the forecasting of salinity in the River Murray at Murray Bridge, South Australia; and the the forecasting of cyanobacteria (Anabaena spp.) in the River Murray at Morgan, South Australia.; Thesis (Ph.D.)--University of Adelaide, School of Civil and Environmental Engineering, 2003; "February 2003."; Corrigenda for, inserted at back; Includes bibliographical references (leaves 475-524 ); xxx, 524 leaves : ill. ; 30 cm.

Artificial neural networks (ANNs), trained to make short term forecasts of algal blooms in lakes and rivers, are potentially useful decision making tools for the operational management of eutrophication. This thesis addresses the question of whether a standardised, gemeric ANN model representation can be developed to achieve this goal. It is argued that four requirements need to be addressed: i) compatibility of models with existing water quality monitoring regimes, ii) stability and repeatability of training outcomes, iii) realistic and meaningful estimates of model performance, and iv) explanation of predictions.; Thesis (Ph.D.)--University of Adelaide, School of Earth and Environmental Sciences, Discipline of Environmental Biology, 2004; "April 2004"; Bibliography: p. 285-299.; xxviii, 299p : ill., map ; 30 cm.; Title page, contents and abstract only. The complete thesis in print form is available from the University Library.

Approved for public release; distribution is unlimited; The clearing of unexploded ordnance (UXO) is a deadly and time consuming process. The U.S. Government is currently spending millions of dollars to remove UXO's from bases that are closing around the world. Existing methods for detecting UXO's only inform the clearing team that a piece of metal is present, rather than the type of metal, either UXO, shrapnel, or garbage. A lot of time and money is spent digging up every piece of metal detected. This thesis presents the use of artificial neural networks to determine the type of UXO that is detected. A multi layered feed forward neural network using the back propagation training algorithm was developed using the language Lisp. The network was trained to recognize five pieces of ammunition. Results from the research show that four out of five pieces of ammunition from the test set were identified with an accuracy of .99 out of 1.0. The network also correctly identified that a tin can was not one of the five pieces of ammunition; http://archive.org/details/usingartificialn00mayj; Captain, United' States Army

This thesis contains the process and results related to optimizing a neural network to predict future positions of airplanes in the vicinity of airports. These predicted positions are then used to calculate future separation distances between pairs of airplanes. The predicted values of the separation distance are used to ensure adequate distances between adjacent aircrafts in the air and, if necessary, to create early warning alarms to alert air traffic control tower personnel about planes that may pass too near each other in the immediate future. The thesis covers three areas of work on this topic. The first section involves optimizing a neural network for Chicago O'Hare Airport. The second is related to gathering data on the performance of this network in different scenarios. These data can be used to determine if the different days/runways have different characteristics. The final phase of this document describes how to generalize the process used to build an optimized neural network for Chicago O'Hare airport in order to provide the capability to easily recreate the process for another airport.; by Geoffrey T. Cooney.; Thesis (M. Eng.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science...

This thesis examines Pulse Coupled Neural Networks (PCNNs) and their applications, and the feasibility of a compact, rugged, cost-efficient optoelectronic implementation. Simulation results are presented. Proposed optical architectures are discussed and analyzed. A new optoelectronic PCNN architecture is also presented. Tradeoffs of optical versus electronic implementations of PCNNs are discussed. This work combines concepts from optical information processing and pulse-coupled neural networks to examine the challenges, limitations, and opportunities of developing an optoelectronic pulse coupled neural network. The analysis finds that, despite advances in optoelectronic technology, fully electronic implementations will still outperform today's proposed optoelectronic implementations in cost, size, flexibility, and ease of implementation.; by Raydiance Wise.; Thesis (S.M.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2007.; Includes bibliographical references (leaves 76-79).

We show that memcapacitive (memory capacitive) systems can be used as synapses in artificial neural networks. As an example of our approach, we discuss the architecture of an integrate-and-fire neural network based on memcapacitive synapses. Moreover, we demonstrate that the spike-timing-dependent plasticity can be simply realized with some of these devices. Memcapacitive synapses are a low-energy alternative to memristive synapses for neuromorphic computation.

A neural network with fixed topology can be regarded as a parametrization of functions, which decides on the correlations between functional variations when parameters are adapted. We propose an analysis, based on a differential geometry point of view, that allows to calculate these correlations. In practise, this describes how one response is unlearned while another is trained. Concerning conventional feed-forward neural networks we find that they generically introduce strong correlations, are predisposed to forgetting, and inappropriate for task decomposition. Perspectives to solve these problems are discussed.; Comment: LaTeX, 7 pages, 3 figures

We present an efficient learning algorithm for the problem of training neural networks with discrete synapses, a well-known hard (NP-complete) discrete optimization problem. The algorithm is a variant of the so-called Max-Sum (MS) algorithm. In particular, we show how, for bounded integer weights with $q$ distinct states and independent concave a priori distribution (e.g. $l_{1}$ regularization), the algorithm's time complexity can be made to scale as $O\left(N\log N\right)$ per node update, thus putting it on par with alternative schemes, such as Belief Propagation (BP), without resorting to approximations. Two special cases are of particular interest: binary synapses $W\in\{-1,1\}$ and ternary synapses $W\in\{-1,0,1\}$ with $l_{0}$ regularization. The algorithm we present performs as well as BP on binary perceptron learning problems, and may be better suited to address the problem on fully-connected two-layer networks, since inherent symmetries in two layer networks are naturally broken using the MS approach.

Despite the widespread practical success of deep learning methods, our theoretical understanding of the dynamics of learning in deep neural networks remains quite sparse. We attempt to bridge the gap between the theory and practice of deep learning by systematically analyzing learning dynamics for the restricted case of deep linear neural networks. Despite the linearity of their input-output map, such networks have nonlinear gradient descent dynamics on weights that change with the addition of each new hidden layer. We show that deep linear networks exhibit nonlinear learning phenomena similar to those seen in simulations of nonlinear networks, including long plateaus followed by rapid transitions to lower error solutions, and faster convergence from greedy unsupervised pretraining initial conditions than from random initial conditions. We provide an analytical description of these phenomena by finding new exact solutions to the nonlinear dynamics of deep learning. Our theoretical analysis also reveals the surprising finding that as the depth of a network approaches infinity, learning speed can nevertheless remain finite: for a special class of initial conditions on the weights, very deep networks incur only a finite, depth independent...

We introduce a class of neural networks derived from probabilistic models in the form of Bayesian networks. By imposing additional assumptions about the nature of the probabilistic models represented in the networks, we derive neural networks with standard dynamics that require no training to determine the synaptic weights, that perform accurate calculation of the mean values of the random variables, that can pool multiple sources of evidence, and that deal cleanly and consistently with inconsistent or contradictory evidence. The presented neural networks capture many properties of Bayesian networks, providing distributed versions of probabilistic models.; Comment: 13 pages, elsarticle

This paper presents an overview of some techniques and concepts coming from dynamical system theory and used for the analysis of dynamical neural networks models. In a first section, we describe the dynamics of the neuron, starting from the Hodgkin-Huxley description, which is somehow the canonical description for the ``biological neuron''. We discuss some models reducing the Hodgkin-Huxley model to a two dimensional dynamical system, keeping one of the main feature of the neuron: its excitability. We present then examples of phase diagram and bifurcation analysis for the Hodgin-Huxley equations. Finally, we end this section by a dynamical system analysis for the nervous flux propagation along the axon. We then consider neuron couplings, with a brief description of synapses, synaptic plasticiy and learning, in a second section. We also briefly discuss the delicate issue of causal action from one neuron to another when complex feedback effects and non linear dynamics are involved. The third section presents the limit of weak coupling and the use of normal forms technics to handle this situation. We consider then several examples of recurrent models with different type of synaptic interactions (symmetric, cooperative, random). We introduce various techniques coming from statistical physics and dynamical systems theory. A last section is devoted to a detailed example of recurrent model where we go in deep in the analysis of the dynamics and discuss the effect of learning on the neuron dynamics. We also present recent methods allowing the analysis of the non linear effects of the neural dynamics on signal propagation and causal action. An appendix...

This thesis investigates a method for using contextual information in text recognition. This is based on the premise that, while reading, humans recognize words with missing or garbled characters by examining the surrounding characters and then selecting the appropriate character. The correct character is chosen based on an inherent knowledge of the language and spelling techniques. We can then model this statistically. The approach taken by this Thesis is to combine feature extraction techniques, Neural Networks and Hidden Markov Modeling. This method of character recognition involves a three step process: pixel image preprocessing, neural network classification and context interpretation. Pixel image preprocessing applies a feature extraction algorithm to original bit mapped images, which produces a feature vector for the original images which are input into a neural network. The neural network performs the initial classification of the characters by producing ten weights, one for each character. The magnitude of the weight is translated into the confidence the network has in each of the choices. The greater the magnitude and separation, the more confident the neural network is of a given choice. The output of the neural network is the input for a context interpreter. The context interpreter uses Hidden Markov Modeling (HMM) techniques to determine the most probable classification for all characters based on the characters that precede that character and character pair statistics. The HMMs are built using an a priori knowledge of the language: a statistical description of the probabilities of digrams. Experimentation and verification of this method combines the development and use of a preprocessor program...

Local learning neural networks have long been limited by their inability to store correlated patterns. A common parameter used to specify the capacity of a network is

Link permanente para citações:

Partly in response to the apparent limitations of explicit symbol processing used by traditional artificial intelligence research, there has been, within the last decade, a growing interest in artificial neural networks. This thesis focuses on the development and testing of a model for describing certain kinds of biological phenomena. The many artificial neural networks available may be classified into three types: (1) self-organizing networks, which have input but no feedback; (2) unsupervised networks, requiring minimal feedback (perhaps a signal indicating success or failure); and (3) supervised models, which employ far more extensive (and, I think, biologically implausible) feedback mechanisms. In this thesis I examine only models of the second type. The Rescorla-Wagner "trial-level" model gives a quantitative description of what happens as a result of a conditioning trial. But that model, along with more detailed, "temporal" (i.e., intratrial) models, such as a traditional Hebbian model and the Sutton-Barto model, make predictions which are at odds with empirical data. Klopf's "drive-reinforcement" model is a much more robust account, from which I develop a simplified drive-reinforcement (SDR) model. I prepare a number of experiments to test my SDR model's correspondence with empirical data derived from animal learning experiments; I demonstrate that the model is capable of describing a wide variety of classical conditioning phenome na; and I 6how how the model may form the basis for instrumental conditioning as well. Finally...

Current methods of speech intelligibility estimation rely on the subjective judgements of trained listeners. Accurate and unbiased intelligibility estimates have a number of procedural and/or methodological constraints including the necessity for large pools of listeners and a wide variety of stimulus materials. Recent research findings however, have shown a strong relationship between speech intelligibility estimates and selected acoustic speech parameters which appear to determine the intelligibility of speech. These findings suggest that such acoustic speech parameters could be used to derive computer-based speech intelligibility estimation, obviating the procedural and methodological constraints typically associated with such estimates. The relationship between speech intelligibility estimates and acoustic speech parameters is complex and nonlinear in nature. Artificial neural networks have proven in general speech recognition that they are capable of dealing with complex and unspecified nonlinear relationships. The purpose of this study was to explore the possibility of using artificial neural networks to make speech intelligibility estimates. Sixty hearing-impaired speakers, whose measured speech intelligibility ranged from 0 to 99%...