This thesis presents some substantial theoretical analyses and optimal treatments of Kohonen's self-organising map (SOM) algorithm, and explores the practical application potential of the algorithm for vector quantisation, pattern classification, and image processing. It consists of two major parts. In the first part, the SOM algorithm is investigated and analysed from a statistical viewpoint. The proof of its universal convergence for any dimensionality is obtained using a novel and extended form of the Central Limit Theorem. Its feature space is shown to be an approximate multivariate Gaussian process, which will eventually converge and form a mapping, which minimises the mean-square distortion between the feature and input spaces. The diminishing effect of the initial states and implicit effects of the learning rate and neighbourhood function on its convergence and ordering are analysed and discussed. Distinct and meaningful definitions, and associated measures, of its ordering are presented in relation to map's fault-tolerance. The SOM algorithm is further enhanced by incorporating a proposed constraint, or Bayesian modification, in order to achieve optimal vector quantisation or pattern classification. The second part of this thesis addresses the task of unsupervised texture-image segmentation by means of SOM networks and model-based descriptions. A brief review of texture analysis in terms of definitions, perceptions, and approaches is given. Markov random field model-based approaches are discussed in detail. Arising from this a hierarchical self-organised segmentation structure, which consists of a local MRF parameter estimator, a SOM network, and a simple voting layer, is proposed and is shown, by theoretical analysis and practical experiment, to achieve a maximum likelihood or maximum a posteriori segmentation. A fast, simple, but efficient boundary relaxation algorithm is proposed as a post-processor to further refine the resulting segmentation. The class number validation problem in a fully unsupervised segmentation is approached by a classical, simple, and on-line minimum mean-square-error method. Experimental results indicate that this method is very efficient for texture segmentation problems. The thesis concludes with some suggestions for further work on SOM neural networks.

A number of researchers have investigated the application of neural networks to visual recognition, with much of the emphasis placed on exploiting the network's ability to generalise. However, despite the benefits of such an approach it is not at all obvious how networks can be developed which are capable of recognising objects subject to changes in rotation, translation and viewpoint. In this study, we suggest that a possible solution to this problem can be found by studying aspects of visual psychology and in particular, perceptual organisation. For example, it appears that grouping together lines based upon perceptually significant features can facilitate viewpoint independent recognition. The work presented here identifies simple grouping measures based on parallelism and connectivity and shows how it is possible to train multi-layer perceptrons (MLPs) to detect and determine the perceptual significance of any group presented. In this way, it is shown how MLPs which are trained via backpropagation to perform individual grouping tasks, can be brought together into a novel, large scale network capable of determining the perceptual significance of the whole input pattern. Finally the applicability of such significance values for recognition is investigated and results indicate that both the NILP and the Kohonen Feature Map can be trained to recognise simple shapes described in terms of perceptual significances. This study has also provided an opportunity to investigate aspects of the backpropagation algorithm, particularly the ability to generalise. In this study we report the results of various generalisation tests. In applying the backpropagation algorithm to certain problems, we found that there was a deficiency in performance with the standard learning algorithm. An improvement in performance could however, be obtained when suitable modifications were made to the algorithm. The modifications and consequent results are reported here.

Images are forming an increasingly large part of modern communications, bringing the need for efficient and effective compression. Many techniques developed for this purpose include transform coding, vector quantization and neural networks. In this thesis, a new neural network method is used to achieve image compression. This work extends the use of 2-layer neural networks to a combination of cascaded networks with one node in the hidden layer. A redistribution of the gray levels in the training phase is implemented in a random fashion to make the minimization of the mean square error applicable to a broad range of images. The computational complexity of this approach is analyzed in terms of overall number of weights and overall convergence. Image quality is measured objectively, using peak signal-to-noise ratio and subjectively, using perception. The effects of different image contents and compression ratios are assessed. Results show the performance superiority of cascaded neural networks compared to that of fixedarchitecture training paradigms especially at high compression ratios. The proposed new method is implemented in MATLAB. The results obtained, such as compression ratio and computing time of the compressed images, are presented.

No description available.

A Dissertation submitted to the Faculty of Science, University of the Witwatersrand, Johannesburg, in fulfilment of the requirements for the degree of Master of Science in Mathematical Statistics,School of Statistics and Actuarial Science. October 2016. / Neural networks (NNs) may be characterised by complex error functions with attributes such as saddle-points, local minima, even-spots and plateaus. This complicates the associated training process in terms of efficiency, convergence and accuracy given that it is done by minimising such complex error functions. This study empirically investigates the performance of two NNs training algorithms which are based on unconstrained and global optimisation theories, i.e. the Resilient propagation (Rprop) and the Conjugate Gradient with Polak-Ribière updates (CGP). It also shows how the network structure plays a role in the training optimisation of NNs. In this regard, various training scenarios are used to classify two protein data, i.e. the Escherichia coli and Yeast data. These training scenarios use varying numbers of hidden nodes and training iterations. The results show that Rprop outperforms CGP. Moreover, it appears that the performance of classifiers varies under various training scenarios. / LG2017

Dentre as inúmeras técnicas de conversão de voz utilizadas atualmente, aquelas baseadas em bancos de filtros wavelet, associadas com redes neurais artificiais,têm se destacado. Este trabalho se concentra em tais técnicas, realizando um estudo que relaciona qual a melhor wavelet para conversão de determinados padrões de voz, apresentando uma análise detalhada de quais são as características que levam a estes resultados. Os testes são realizados com vozes da base de dados TIMIT do Linguistic Data Consortium (LDC). / Dentre as inúmeras técnicas de conversão de voz utilizadas atualmente, aquelas baseadas em bancos de filtros wavelet, associadas com redes neurais artificiais, têm se destacado. Este trabalho se concentra em tais técnicas, realizando um estudo que relaciona qual a melhor wavelet para conversão de determinados padrões de voz, apresentando uma análise detalhada de quais são as características que levam a estes resultados. Os testes são realizados com vozes da base de dados TIMIT do Linguistic Data Consortium (LDC).

This thesis considers the neural network learning control of a variable-geometry automotive active suspension system which combines most of the benefits of active suspension systems with low energy consumption. Firstly, neural networks are applied to the control of various simplified automotive active suspensions, in order to understand how a neural network controller can be integrated with a physical dynamic system model. In each case considered, the controlled system has a defined objective and the minimisation of a cost function. The neural network is set up in a learning structure, such that it systematically improves the system performance via repeated trials and modifications of parameters. The learning efficiency is demonstrated by the given system performance in agreement with prior results for both linear and non-linear systems. The above simulation results are generated by MATLAB and the Neural Network Toolbox. Secondly, a half-car model, having one axle and an actuator on each side, is developed via the computer language, AUTOSIM. Each actuator varies the ratio of the spring/damper unit length change to wheel displacement in order to control each wheel rate. The neural network controller is joined with the half-car model and learns to reduce the defined cost function containing a weighted sum of the squares of the body height change, body roll and actuator displacements. The performances of the neurocontrolled system are compared with those of passive and proportional-plusdifferential controlled systems under various conditions. These involve various levels of lateral force inputs and vehicle body weight changes. Finally, energy consumption of the variable-geometry system, with either the neurocontrol or proportional-plus-differential control, is analysed using an actuator model via the computer simulation package, SIMULINK. The simulation results are compared with those of other actively-controlled suspension systems taken from the literature.

Previous research that has evaluated the effectiveness of personality variables for predicting work performance has predominantly relied on methods designed to detect simple relationships. The research reported in this thesis employed artificial neural networks ??? a method that is capable of capturing complex nonlinear and configural relationships among variables ??? and the findings were compared to those obtained by the more traditional method of linear regression. Six datasets that comprise a range of occupations, personality inventories, and work performance measures were used as the basis of the analyses. A series of studies were conducted to compare the predictive performance of prediction equations that a) were developed using either artificial neural networks or linear regression, and b) differed with respect to the type and number of personality variables that were used as predictors of work performance. Studies 1 and 2 compared the two methods using individual personality variables that assess the broad constructs of the five-factor model of personality. Studies 3 and 4 used combinations of these broad variables as the predictors. Study 5 employed narrow personality variables that assess specific facets of the broad constructs. Additional methodological contributions include the use of a resampling procedure, the use of multiple measures of predictive performance, and the comparison of two procedures for developing neural networks. Across the studies, it was generally found that the neural networks were rarely able to outperform the simpler linear regression equations, and this was attributed to the lack of reliable nonlinearity and configurality in personality-work performance relationships. However, the neural networks were able to outperform linear regression in the few instances where there was some independent evidence of nonlinear or configural relationships. Consequently, although the findings do not support the usefulness of neural networks for specifically improving the effectiveness of personality variables as predictors of work performance, in a broader sense they provide some grounds for optimism for organisational researchers interested in applying this method to investigate and exploit complex relationships among variables.

This thesis explores the relationship between two classification models: decision trees and multilayer perceptrons. Decision trees carve up databases into box-shaped regions, and make predictions based on the majority class in each box. They are quick to build and relatively easy to interpret. Multilayer perceptrons (MLPs) are often more accurate than decision trees, because they are able to use soft, curved, arbitrarily oriented decision boundaries. Unfortunately MLPs typically require a great deal of effort to determine a good number and arrangement of neural units, and then require many passes through the database to determine a good set of connection weights. The cost of creating and training an MLP is thus hundreds of times greater than the cost of creating a decision tree, for perhaps only a small gain in accuracy. The following scheme is proposed for reducing the computational cost of creating and training MLPs. First, build and prune a decision tree to generate prior knowledge of the database. Then, use that knowledge to determine the initial architecture and connection weights of an MLP. Finally, use a training algorithm to refine the knowledge now embedded in the MLP. This scheme has two potential advantages: a suitable neural network architecture is determined very quickly, and training should require far fewer passes through the data. In this thesis, new algorithms for initialising MLPs from decision trees are developed. The algorithms require just one traversal of a decision tree, and produce four-layer MLPs with the same number of hidden units as there are nodes in the tree. The resulting MLPs can be shown to reach a state more accurate than the decision trees that initialised them, in fewer training epochs than a standard MLP. Employing this approach typically results in MLPs that are just as accurate as standard MLPs, and an order of magnitude cheaper to train.

<p>In this work new artificial learning and innate control mechanisms are proposed for application</p><p>in autonomous behavioral systems for mobile robots. An autonomous system (for mobile robots)</p><p>existent in the literature is enhanced with respect to its capacity of exploring the environment and</p><p>avoiding risky configurations (that lead to collisions with obstacles even after learning). The</p><p>particular autonomous system is based on modular hierarchical neural networks. Initially,the</p><p>autonomous system does not have any knowledge suitable for exploring the environment (and</p><p>capture targets œ foraging). After a period of learning,the system generates efficientobstacle</p><p>avoid ance and target seeking behaviors. Two particular deficiencies of the forme rautonomous</p><p>system (tendency to generate unsuitable cyclic trajectories and ineffectiveness in risky</p><p>configurations) are discussed and the new learning and controltechniques (applied to the</p><p>autonomous system) are verified through simulations. It is shown the effectiveness of the</p><p>proposals: theautonomous system is able to detect unsuitable behaviors (cyclic trajectories) and</p><p>decrease their probability of appearance in the future and the number of collisions in risky</p><p>situations is significantly decreased. Experiments also consider maze environments (with targets</p><p>distant from each other) and dynamic environments (with moving objects).</p>