A new approach is proposed which uses a combination of a Backprop paradigm neural network along with some perceptron processing elements performing logic operations to construct a numeric-to-symbolic converter. The design approach proposed herein is capable of implementing a decision region defined by a multi-dimensional, non-linear boundary surface. By defining a "two-valued" subspace of the boundary surface, a Backprop paradigm neural network is used to model the boundary surf ace. An input vector is tested by the neural network boundary model (along with perceptron logic gates) to determine whether the incoming vector point is within the decision region or not. Experiments with two qualitatively different kinds of nonlinear surface were carried out to test and demonstrate the design approach.

La classification est l'attribution d'une classe spécifique à un objet donné. Cette attribution a besoin d'un certain degré d'abstraction pour pouvoir extraire des généralités à partir des exemples dont on dispose. Pour une machine, la classification de visages, de données médicales, de formes, sont toutes des tâches assez difficiles. Par exemple, dans le cas de la reconnaissance de caractères manuscrits, il est difficile d'énoncer une description générale qui tienne compte de toutes les variations particulières de chaque caractère. Une autre approche qui peut être utilisée pour cette tâche est celle de l'apprentissage. Ainsi, le critère pour décider si une image correspond ou non à une lettre 'A' consiste à comparer si cette image est (suffisamment similaire} à des 'a' vus auparavant. De ce point de vue, on ne calcule pas la classification de caractères : elle doit être apprise à partir d'exemples. Ces dernières années, de nouvelles techniques neuronales d'apprentissage ont été développées. Cet apprentissage avec des réseaux de neurones se fait actuellement en suivant deux approches : certains algorithmes comme la Rétropropagation du Gradient ont besoin d'introduire a priori le nombre et la connectivité des unités cachées et déterminer les poids des connexions par minimisation d'un coût. Le réseau ainsi obtenu est éventuellement élagué. Avec une approche constructive on apprend en même temps le nombre d'unités et les poids, dans le cadre d'une architecture fixée, commençant généralement avec une seule unité. Le but de cette thèse est de présenter de nouvelles heuristiques pour générer, d'une manière constructive, des réseaux de neurones pour la classification. Elles permettent de générer des réseaux à une seule couche cachée complètement connectée aux unités d'entrée, et un neurone de sortie connecté aux unités cachées. Les neurones cachés et de sortie sont des unités binaires, pouvant faire soit des séparations linéaires, soit des séparations sphériques. Ces heuristiques sont couplées avec des algorithmes d'apprentissage pour le perceptron, Minimerror-L pour les séparations linéaires et Minimerror-S pour les séparations sphériques. Trois nouveaux algorithmes constructifs, qui différent suivant le type de neurones cachés et aussi suivant la définition des cibles que ceux-ci doivent apprendre on été développés. Pendant le processus d'apprentissage, des neurones cachés entraînés pour apprendre ces cibles vont diminuer le nombre d'erreurs de classification du neurone de sortie. Les réseaux ainsi bâtis ont généralement moins de paramètres (poids) et généralisent mieux que les réseaux entraînés avec d'autres algorithmes.

In this research, a novel machine learning method is designed and applied to Mine Countermeasure Missions. Similarly to some kernel methods, the proposed approach seeks to compute a linear model from another higher dimensional feature space. However, no kernel is used and the feature mapping is explicit. Computation can be done directly in the accessible feature space. In the proposed approach, the feature projection is implemented by constructing a large hidden layer, which differs from traditional belief that Multi-Layer Perceptron is usually funnel-shaped and the hidden layer is used as feature extractor. The proposed approach is a general method that can be applied to various problems. It is able to improve the performance of the neural network based methods and the learning speed of support vector machine. The classification speed of the proposed approach is also faster than that of kernel machines on the mine countermeasure mission task.

No description available.

Digital images acquired at construction sites contain valuable information useful for various applications including As-built documentation of building elements, effective progress monitoring, structural damage assessment, and quality control of construction material. As a result there is an increasing need for effective methods to recognize different building materials in digital images and videos. Pattern recognition is a mature field within the area of image processing; however, its application in the area of civil engineering and building construction is only recent. In order to develop any robust image recognition method, it is necessary to choose the optimal machine learning algorithm. To generate a robust color model for building material detection in an outdoor construction environment, a comparative analysis of three generative and discriminative machine learning algorithms, namely, multilayer perceptron (MLP), radial basis function (RBF), and support vector machines (SVMs), is conducted. The main focus of this study is on three classes of building materials: concrete, plywood, and brick. For training purposes a large-size data set including hundreds of images is collected. The comparison study is conducted by implementing necessary algorithms in MATLAB and testing over hundreds of construction-site images. To evaluate the performance of each technique, the results are compared with a manual classification of building materials. In order to better assess the performance of each technique, experiments are conducted by taking pictures under various realistic jobsite conditions, e.g., different ranges of image resolutions, different distance of camera from object, and different types of cameras.

Made available in DSpace on 2016-06-02T19:05:16Z (GMT). No. of bitstreams: 1 DissLGPN.pdf: 973893 bytes, checksum: d6c9825733402bdd76187d2779e3f8b0 (MD5) Previous issue date: 2004-02-27 / Universidade Federal de Sao Carlos / Constructive neural learning is a neural learning model that does not assume a fixed topology before training begins. The main characteristic of this learning model is the dynamic construction of the network hidden layers which occurs simultaneously with training. This research work investigates six constructive neural algorithms namely, tower, pyramid, tiling, upstart, Distal and cascade-correlation, evaluating each of them with relation to advantages and disadvantages, ease of training, size and topology of the network, restrictions and performance. The work presents a computational system (CONEB) which implements each algorithm. Results obtained by using the different algorithms in several knowledge domains are presented and analysed. / Aprendizado Neural Construtivo é um modelo de aprendizado neural que não pressupõe a definição de uma topologia de rede fixa antes do início do treinamento. A principal característica deste modelo de aprendizado é a construção dinâmica das camadas intermediárias da rede, à medida que vão sendo necessárias ao seu treinamento. Este trabalho de pesquisa investiga seis algoritmos neurais construtivos, a saber, tower, pyramid, tiling, upstart, Distal e cascade-correlation, buscando avaliar cada um deles com relação a vantagens e desvantagens, facilidade no treinamento, tamanho e topologia de rede criada, restrições de uso e desempenho. O trabalho apresenta um ambiente computacional (CONEB) que disponibiliza a implementação de cada um dos algoritmos. São apresentados e analisados os resultados obtidos utilizando os diferentes algoritmos em vários domínios de conhecimento.

Interaction of various components in rotating machinery like gearboxes may generate excitation forces at various frequencies. These frequencies may sometimes overlap with the frequencies of the forces generated by other components in the system. Conventional vibration spectrum analysis does not attenuate noise and spectral frequency band overlapping, which in many applications masks the changes in the structural response caused by the deterioration of certain components in the machine. This problem is overcome by the use of time domain averaging (dsynchronous averaging). In time domain averaging, the vibration signal is sampled at a frequency that is synchronized with the rotation of the gear of interest and the samples obtained for each singular position of the gear are ensemble-averaged. When sufficient averages are taken, all the vibration from the gearbox, which is asynchronous with the vibration of the gear, is attenuated. The resulting time synchronously averaged signal obtained through this process indicates the vibration produced during one rotation of the monitored gear. This direct time domain averaging process essentially acts as a broadband noise synchronous filter, which filters out the frequency content that is asynchronous with the vibration of the gear of interest provided that enough averages are taken. The time domain averaging procedure requires an enormous amount of vibration data to execute, making it very difficult to develop online gearbox condition monitoring systems that make use of time domain averaging to enhance their diagnostic capabilities since data acquisition and analysis cannot be done simultaneously. The objective of this research was to develop a more efficient way for calculating the time domain average of a gear vibration signal. A study of Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) was conducted to assess their suitability for use in time domain averaging. Two time domain averaging models that use ANNs and SVMs were developed. Model 1 uses a single feedforward network configuration to map the input which are rotation synchronized gear vibration signals to the output which is the time domain average of the gear vibration signal, using only a section of the input space. Model 2 operates in two stages. In the first stage, it uses a feedforward network to predict the instantaneous time domain average of the gear vibration after 10 inputs (10 rotation synchronized gear vibration signals) to predict the instantaneous average of the gear rotation. The outputs from the first state are used as inputs to the second stage, where a second feedforward network is used to predict the time domain average of the entire vibration signal. When ANNs and SVMs were implemented, the results indicated that the amount of gear vibration date that is required to calculate the time domain average using Model 1 can be reduced by 75 percent and the amount of gear vibration data that needs to be stored in the data acquisition system when Model 2 is used can be reduced by 83 percent. / Dissertation (M Eng (Mechanical and Aeronautical Engineering))--University of Pretoria, 2006. / Mechanical and Aeronautical Engineering / unrestricted

In this research, a novel machine learning method is designed and applied to Mine Countermeasure Missions. Similarly to some kernel methods, the proposed approach seeks to compute a linear model from another higher dimensional feature space. However, no kernel is used and the feature mapping is explicit. Computation can be done directly in the accessible feature space. In the proposed approach, the feature projection is implemented by constructing a large hidden layer, which differs from traditional belief that Multi-Layer Perceptron is usually funnel-shaped and the hidden layer is used as feature extractor. The proposed approach is a general method that can be applied to various problems. It is able to improve the performance of the neural network based methods and the learning speed of support vector machine. The classification speed of the proposed approach is also faster than that of kernel machines on the mine countermeasure mission task.

No description available.

L’apprentissage automatique fait parti d’une branche de l’intelligence artificielle et est utilisé dans de nombreux domaines en science. Il se divise en trois catégories principales : supervisé, non-supervisé et par renforcement. Ce mémoire de maîtrise portera uniquement sur l’apprentissage supervisé et plus précisément sur la classification de données. Un des premiers algorithmes en classification, le perceptron, fut proposé dans les années soixante. Nous proposons une variante de cet algorithme, que nous appelons le perceptron dual tronqué, qui permet l’arrêt de l’algorithme selon un nouveau critère. Nous comparerons cette nouvelle variante à d’autres variantes du perceptron. De plus, nous utiliserons le perceptron dual tronqué pour construire des classificateurs plus complexes comme les «Bayes Point Machines». / Machine Learning is a part of the artificial intelligence and is used in many fields in science. It is divided into three categories : supervised, not supervised and by reinforcement. This master’s paper will relate only the supervised learning and more precisely the classification of datas. One of the first algorithms in classification, the perceptron, was proposed in the Sixties. We propose an alternative of this algorithm, which we call the truncated dual perceptron, which allows the stop of the algorithm according to a new criterion. We will compare this new alternative with other alternatives of the perceptron. Moreover, we will use the truncated dual perceptron to build more complex classifiers like the «Bayes Point Machines».