Support Vector Machines: Training and Applications

The Support Vector Machine (SVM) is a new and very promising classification technique developed by Vapnik and his group at AT&T Bell Labs. This new learning algorithm can be seen as an alternative training technique for Polynomial, Radial Basis Function and Multi-Layer Perceptron classifiers. An interesting property of this approach is that it is an approximate implementation of the Structural Risk Minimization (SRM) induction principle. The derivation of Support Vector Machines, its relationship with SRM, and its geometrical insight, are discussed in this paper. Training a SVM is equivalent to solve a quadratic programming problem with linear and box constraints in a number of variables equal to the number of data points. When the number of data points exceeds few thousands the problem is very challenging, because the quadratic form is completely dense, so the memory needed to store the problem grows with the square of the number of data points. Therefore, training problems arising in some real applications with large data sets are impossible to load into memory, and cannot be solved using standard non-linear constrained optimization algorithms. We present a decomposition algorithm that can be used to train SVM's over large data sets. The main idea behind the decomposition is the iterative solution of sub-problems and the evaluation of, and also establish the stopping criteria for the algorithm. We present previous approaches, as well as results and important details of our implementation of the algorithm using a second-order variant of the Reduced Gradient Method as the solver of the sub-problems. As an application of SVM's, we present preliminary results we obtained applying SVM to the problem of detecting frontal human faces in real images.

Statistical downscaling of river flows

An extensive statistical ‘downscaling’ study is done to relate large-scale climate information from a general circulation model (GCM) to local-scale river flows in SW France for 51 gauging stations ranging from nival (snow-dominated) to pluvial (rainfall-dominated) river-systems. This study helps to select the appropriate statistical method at a given spatial and temporal scale to downscale hydrology for future climate change impact assessment of hydrological resources. The four proposed statistical downscaling models use large-scale predictors (derived from climate model outputs or reanalysis data) that characterize precipitation and evaporation processes in the hydrological cycle to estimate summary flow statistics. The four statistical models used are generalized linear (GLM) and additive (GAM) models, aggregated boosted trees (ABT) and multi-layer perceptron neural networks (ANN). These four models were each applied at two different spatial scales, namely at that of a single flow-gauging station (local downscaling) and that of a group of flow-gauging stations having the same hydrological behaviour (regional downscaling). For each statistical model and each spatial resolution, three temporal resolutions were considered, namely the daily mean flows, the summary statistics of fortnightly flows and a daily ‘integrated approach’. The results show that flow sensitivity to atmospheric factors is significantly different between nival and pluvial hydrological systems which are mainly influenced, respectively, by shortwave solar radiations and atmospheric temperature. The non-linear models (i.e. GAM, ABT and ANN) performed better than the linear GLM when simulating fortnightly flow percentiles. The aggregated boosted trees method showed higher and less variable R2 values to downscale the hydrological variability in both nival and pluvial regimes. Based on GCM cnrm-cm3 and scenarios A2 and A1B, future relative changes of fortnightly median flows were projected based on the regional downscaling approach. The results suggest a global decrease of flow in both pluvial and nival regimes, especially in spring, summer and autumn, whatever the considered scenario. The discussion considers the performance of each statistical method for downscaling flow at different spatial and temporal scales as well as the relationship between atmospheric processes and flow variability.

Neural and wavelet network models for financial distress classification

This work analyzes the use of linear discriminant models, multi-layer perceptron neural networks and wavelet networks for corporate financial distress prediction. Although simple and easy to interpret, linear models require statistical assumptions that may be unrealistic. Neural networks are able to discriminate patterns that are not linearly separable, but the large number of parameters involved in a neural model often causes generalization problems. Wavelet networks are classification models that implement nonlinear discriminant surfaces as the superposition of dilated and translated versions of a single "mother wavelet" function. In this paper, an algorithm is proposed to select dilation and translation parameters that yield a wavelet network classifier with good parsimony characteristics. The models are compared in a case study involving failed and continuing British firms in the period 1997-2000. Problems associated with over-parameterized neural networks are illustrated and the Optimal Brain Damage pruning technique is employed to obtain a parsimonious neural model. The results, supported by a re-sampling study, show that both neural and wavelet networks may be a valid alternative to classical linear discriminant models.

Comparison of extrasystolic ECG signal classifiers using discrete wavelet transforms

This work compares and contrasts results of classifying time-domain ECG signals with pathological conditions taken from the MITBIH arrhythmia database. Linear discriminant analysis and a multi-layer perceptron were used as classifiers. The neural network was trained by two different methods, namely back-propagation and a genetic algorithm. Converting the time-domain signal into the wavelet domain reduced the dimensionality of the problem at least 10-fold. This was achieved using wavelets from the db6 family as well as using adaptive wavelets generated using two different strategies. The wavelet transforms used in this study were limited to two decomposition levels. A neural network with evolved weights proved to be the best classifier with a maximum of 99.6% accuracy when optimised wavelet-transform ECG data wits presented to its input and 95.9% accuracy when the signals presented to its input were decomposed using db6 wavelets. The linear discriminant analysis achieved a maximum classification accuracy of 95.7% when presented with optimised and 95.5% with db6 wavelet coefficients. It is shown that the much simpler signal representation of a few wavelet coefficients obtained through an optimised discrete wavelet transform facilitates the classification of non-stationary time-variant signals task considerably. In addition, the results indicate that wavelet optimisation may improve the classification ability of a neural network. (c) 2005 Elsevier B.V. All rights reserved.

Neural networks for linear control: an analysis

This paper considers the use of radial basis function and multi-layer perceptron networks for linear or linearizable, adaptive feedback control schemes in a discrete-time environment. A close look is taken at the model structure selected and the extent of the resulting parameterization. A comparison is made with standard, nonneural network algorithms, e.g. self-tuning control.

A critique of neural networks for discrete-time linear control

This paper discusses the use of multi-layer perceptron networks for linear or linearizable, adaptive feedback.control schemes in a discrete-time environment. A close look is taken at the model structure selected and the extent of the resulting parametrization. A comparison is made with standard, non-perceptron algorithms, e.g. self-tuning control, and it is shown how gross over-parametrization can occur in the neural network case. Because of the resultant heavy computational burden and poor controller convergence, a strong case is made against the use of neural networks for discrete-time linear control.

The control of dynamical systems by neural networks

In this paper the use of neural networks for the control of dynamical systems is considered. Both identification and feedback control aspects are discussed as well as the types of system for which neural networks can provide a useful technique. Multi-layer Perceptron and Radial Basis function neural network types are looked at, with an emphasis on the latter. It is shown how basis function centre selection is a critical part of the implementation process and that multivariate clustering algorithms can be an extremely useful tool for finding centres.

Measuring the influence of fibre-to-fabric properties on the pilling of wool fabrics

This study ranks the contribution of various fibre, yarn and fabric attributes to the pilling of wool knitwear. On the basis of an artificial neural network modelling, a combination of sensitivity analysis, forwards/backwards search and genetic algorithms was used to identify the importance of various fibre/yarn/fabric input parameters. The three different techniques show broad similarities in their assessment of which input parameters are important or are not important in affecting fabric pilling. The ranking shows that fabric cover factor has the most effect on pilling, followed by yarn count and thin places, fibre length, yarn twist, etc. It is further illustrated that the directional trend of the predicted pilling outputs for a selection of inputs was in line with the expected behaviour. To verify the findings of input feature selection, input factors deemed to have a small effect on the predicted pilling output, such as fibre length and diameter variations and curvature, were removed and the subsequent performance statistically compared to the original multi-layer perceptron. Differences between the outputs predicted by the original and pruned models are found not to be statistically significant at the 5% significance level. Results from this study may help manufacturers and knitwear designers in choosing the most appropriate materials and structures to reduce the pilling propensity of wool knitwear.

Detection and classification of road signs in natural environments

An automatic road sign recognition system first locates road signs within images captured by an imaging sensor on-board of a vehicle, and then identifies the detected road signs. This paper presents an automatic neural-network-based road sign recognition system. First, a study of the existing road sign recognition research is presented. In this study, the issues associated with automatic road sign recognition are described, the existing methods developed to tackle the road sign recognition problem are reviewed, and a comparison of the features of these methods is given. Second, the developed road sign recognition system is described. The system is capable of analysing live colour road scene images, detecting multiple road signs within each image, and classifying the type of road signs detected. The system consists of two modules: detection and classification. The detection module segments the input image in the hue-saturation-intensity colour space, and then detects road signs using a Multi-layer Perceptron neural-network. The classification module determines the type of detected road signs using a series of one to one architectural Multi-layer Perceptron neural networks. Two sets of classifiers are trained using the Resillient-Backpropagation and Scaled-Conjugate-Gradient algorithms. The two modules of the system are evaluated individually first. Then the system is tested as a whole. The experimental results demonstrate that the system is capable of achieving an average recognition hit-rate of 95.96% using the scaled-conjugate-gradient trained classifiers.

A hybrid neural classifier for dimensionality reduction and data visualization and its application to fault detection and classification of induction motors

In this paper, a hybrid neural classifier combining the auto-encoder neural network and the Lattice Vector Quantization (LVQ) model is described. The auto-encoder network is used for dimensionality reduction by projecting high dimensional data into the 2D space. The LVQ model is used for data visualization by forming and adapting the granularity of a data map. The mapped data are employed to predict the target classes of new data samples. To improve classification accuracy, a majority voting scheme is adopted by the hybrid classifier. To demonstrate the applicability of the hybrid classifier, a series of experiments using simulated and real fault data from induction motors is conducted. The results show that the hybrid classifier is able to outperform the Multi-Layer Perceptron neural network, and to produce very good classification accuracy rates for various fault conditions of induction motors.

Data mining analysis of an urban tunnel pressure drop based on CFD data

An accurate estimation of pressure drop due to vehicles inside an urban tunnel plays a pivotal role in tunnel ventilation issue. The main aim of the present study is to utilize computational intelligence technique for predicting pressure drop due to cars in traffic congestion in urban tunnels. A supervised feed forward back propagation neural network is utilized to estimate this pressure drop. The performance of the proposed network structure is examined on the dataset achieved from Computational Fluid Dynamic (CFD) simulation. The input data includes 2 variables, tunnel velocity and tunnel length, which are to be imported to the corresponding algorithm in order to predict presure drop. 10-fold Cross validation technique is utilized for three data mining methods, namely: multi-layer perceptron algorithm, support vector machine regression, and linear regression. A comparison is to be made to show the most accurate results. Simulation results illustrate that the Multi-layer perceptron algorithm is able to accurately estimate the pressure drop.

Classificação de imagens de sensoriamento remoto baseada em textura por redes neurais

O objetivo principal deste trabalho é propor uma metodologia de classificação de imagens de sensoriamento remoto que integre a importância de atributos de textura na seleção de feições, através da utilização de freqüências espaciais de cada classe textural e sua direção, com a eficiência das redes neurais artificiais para classificá-las. O processo é composto por uma etapa de filtragem baseada nos filtros de Gabor, seguida de uma fase de classificação através de uma rede neural Multi-Layer Perceptron com algoritmo BackPropagation. A partir da transformada de Fourier são estimados os parâmetros a serem utilizados na constituição dos filtros de Gabor, adequados às freqüências espaciais associadas a cada classe presente na imagem a ser classificada. Desta forma, cada filtro gera uma imagem filtrada. O conjunto de filtros determina um conjunto de imagens filtradas (canais texturais). A classificação pixel a pixel é realizada pela rede neural onde cada pixel é definido por um vetor de dimensionalidade igual ao número de filtros do conjunto. O processo de classificação através da rede neural Multi-Layer Perceptron foi realizado pelo método de classificação supervisionada. A metodologia de classificação de imagens de sensoriamento remoto proposta neste trabalho foi testada em imagens sintética e real de dimensões 256 x 256 pixels. A análise dos resultados obtidos é apresentada sob a forma de uma Matriz de Erros, juntamente com a discussão dos mesmos.

Wear Particle Classifier System Based on an Artificial Neural Network

This paper describes a method of identifying morphological attributes that classify wear particles in relation to the wear process from which they originate and permit the automatic identification without human expertise. The method is based on the use of Multi Layer Perceptron (MLP) for analysis of specific types of microscopic wear particles. The classification of the wear particles was performed according to their morphological attributes of size and aspect ratio, among others. (C) 2010 Journal of Mechanical Engineering. All rights reserved.

Artificial neural networks: A novel approach to analysing the nutritional ecology of a blowfly species, Chrysomya megacephala

Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP)