Search

In multiple instance learning problems, patterns are often given as bags and each bag consists of some instances. Most of existing research in the area focuses on multiple instance classification and multiple instance regression, while very limited work has been conducted for multiple instance clustering (MIC). This paper formulates a novel framework, maximum margin multiple instance clustering (M(3)IC), for MIC. However, it is impractical to directly solve the optimization problem of M(3)IC. Therefore, M(3)IC is relaxed in this paper to enable an efficient optimization solution with a combination of the constrained concave-convex procedure and the cutting plane method. Furthermore, this paper presents some important properties of the proposed method and discusses the relationship between the proposed method and some other related ones. An extensive set of empirical results are shown to demonstrate the advantages of the proposed method against existing research for both effectiveness and efficiency.

This paper introduces the Clifford support vector machines (CSVM) as a generalization of the real and complex-valued support vector machines using the Clifford geometric algebra. In this framework, we handle the design of kernels involving the Clifford or geometric product. In this approach, one redefines the optimization variables as multivectors. This allows us to have a multivector as output. Therefore, we can represent multiple classes according to the dimension of the geometric algebra in which we work. We show that one can apply CSVM for classification and regression and also to build a recurrent CSVM. The CSVM is an attractive approach for the multiple input multiple output processing of high-dimensional geometric entities. We carried out comparisons between CSVM and the current approaches to solve multiclass classification and regression. We also study the performance of the recurrent CSVM with experiments involving time series. The authors believe that this paper can be of great use for researchers and practitioners interested in multiclass hypercomplex computing, particularly for applications in complex and quaternion signal and image processing, satellite control, neurocomputation, pattern recognition, computer vision, augmented virtual reality, robotics, and humanoids.

Soft computing (SC) emerged as an integrating framework for a number of techniques that could complement one another quite well (artificial neural networks, fuzzy systems, evolutionary algorithms, probabilistic reasoning). Since its inception, a distinctive goal has been to dig out the deep relationships among their components. This paper considers two wide families of SC models. On the one hand, the regime-switching autoregressive paradigm is a recent development in statistical time series modeling, and it includes a set of models closely related to artificial neural networks. On the other hand, we consider fuzzy rule-based systems in the framework of time series analysis. This paper discloses original results establishing functional equivalences between models of these two classes, and hence opens the door to a productive line of research where results and techniques from one area can be applied in the other. As a consequence of the equivalences presented in this paper, we prove the asymptotic stationarity of a class of fuzzy rule-based systems. Simulations based on information criteria show the importance of the selection of the proper membership function.

A crucial issue in designing learning machines is to select the correct model parameters. When the number of available samples is small, theoretical sample-based generalization bounds can prove effective, provided that they are tight and track the validation error correctly. The maximal discrepancy (MD) approach is a very promising technique for model selection for support vector machines (SVM), and estimates a classifier's generalization performance by multiple training cycles on random labeled data. This paper presents a general method to compute the generalization bounds for SVMs, which is based on referring the SVM parameters to an unsupervised solution, and shows that such an approach yields tight bounds and attains effective model selection. When one estimates the generalization error, one uses an unsupervised reference to constrain the complexity of the learning machine, thereby possibly decreasing sharply the number of admissible hypothesis. Although the methodology has a general value, the method described in the paper adopts vector quantization (VQ) as a representation paradigm, and introduces a biased regularization approach in bound computation and learning. Experimental results validate the proposed method on complex real-world data sets.

In this paper, a sparse learning algorithm, probabilistic classification vector machines (PCVMs), is proposed. We analyze relevance vector machines (RVMs) for classification problems and observe that adopting the same prior for different classes may lead to unstable solutions. In order to tackle this problem, a signed and truncated Gaussian prior is adopted over every weight in PCVMs, where the sign of prior is determined by the class label, i.e., +1 or -1. The truncated Gaussian prior not only restricts the sign of weights but also leads to a sparse estimation of weight vectors, and thus controls the complexity of the model. In PCVMs, the kernel parameters can be optimized simultaneously within the training algorithm. The performance of PCVMs is extensively evaluated on four synthetic data sets and 13 benchmark data sets using three performance metrics, error rate (ERR), area under the curve of receiver operating characteristic (AUC), and root mean squared error (RMSE). We compare PCVMs with soft-margin support vector machines (SVM(Soft)), hard-margin support vector machines (SVM(Hard)), SVM with the kernel parameters optimized by PCVMs (SVM(PCVM)), relevance vector machines (RVMs), and some other baseline classifiers. Through five replications of twofold cross-validation F test, i.e., 5 x 2 cross-validation F test, over single data sets and Friedman test with the corresponding post-hoc test to compare these algorithms over multiple data sets, we notice that PCVMs outperform other algorithms, including SVM(Soft), SVM(Hard), RVM, and SVM(PCVM), on most of the data sets under the three metrics, especially under AUC. Our results also reveal that the performance of SVM(PCVM) is slightly better than SVM(Soft), implying that the parameter optimization algorithm in PCVMs is better than cross validation in terms of performance and computational complexity. In this paper, we also discuss the superiority of PCVMs' formulation using maximum a posteriori (MAP) analysis and margin analysis, which explain the empirical success of PCVMs.

The self-organizing network (SON)-based multiple model system is a recently proposed method for identifying the dynamics of a general nonlinear system. It has been observed by researchers that cooperative learning among neighboring regions is sometimes important for the success of identification of a nonlinear system under the multiple model system framework. In this paper, we intend to formally evaluate the effects of cooperative learning and topology preservation in identification of multiple model system based on SON. The results of the mathematical analysis supports the heuristic that a good learning strategy for identifying the local model parameters of a SON-based multiple model system is to choose a neighborhood function whose effective region is initially wider and is reduced gradually during learning. An example of nonlinear function approximation is also provided at the end of this paper to demonstrate the results of the mathematical analysis.

As is well known in statistics, the resulting linear regressors by using the rank-based Wilcoxon approach to linear regression problems are usually robust against (or insensitive to) outliers. This motivates us to introduce in this paper the Wilcoxon approach to the area of machine learning. Specifically, we investigate four new learning machines, namely Wilcoxon neural network (WNN), Wilcoxon generalized radial basis function network (WGRBFN), Wilcoxon fuzzy neural network (WFNN), and kernel-based Wilcoxon regressor (KWR). These provide alternative learning machines when faced with general nonlinear learning problems. Simple weights updating rules based on gradient descent will be derived. Some numerical examples will be provided to compare the robustness against outliers for various learning machines. Simulation results show that the Wilcoxon learning machines proposed in this paper have good robustness against outliers. We firmly believe that the Wilcoxon approach will provide a promising methodology for many machine learning problems.

This paper is concerned with an enhanced independent component analysis (ICA) and its application to face recognition. Typically, face representations obtained by ICA involve unsupervised learning and high-order statistics. In this paper, we develop an enhancement of the generic ICA by augmenting this method by the Fisher linear discriminant analysis (LDA); hence, its abbreviation, FICA. The FICA is systematically developed and presented along with its underlying architecture. A comparative analysis explores four distance metrics, as well as classification with support vector machines (SVMs). We demonstrate that the FICA approach leads to the formation of well-separated classes in low-dimension subspace and is endowed with a great deal of insensitivity to large variation in illumination and facial expression. The comprehensive experiments are completed for the facial-recognition technology (FERET) face database; a comparative analysis demonstrates that FICA comes with improved classification rates when compared with some other conventional approaches such as eigenface, fisherface, and the ICA itself.

In this paper, we propose a new neural architecture for classification of visual patterns that is motivated by the two concepts of image pyramids and local receptive fields. The new architecture, called pyramidal neural network (PyraNet), has a hierarchical structure with two types of processing layers: Pyramidal layers and one-dimensional (1-D) layers. In the new network, nonlinear two-dimensional (2-D) neurons are trained to perform both image feature extraction and dimensionality reduction. We present and analyze five training methods for PyraNet [gradient descent (GD), gradient descent with momentum, resilient back-propagation (RPROP), Polak-Ribiere conjugate gradient (CG), and Levenberg-Marquadrt (LM)] and two choices of error functions [mean-square-error (mse) and cross-entropy (CE)]. In this paper, we apply PyraNet to determine gender from a facial image, and compare its performance on the standard facial recognition technology (FERET) database with three classifiers: The convolutional neural network (NN), the k-nearest neighbor (k-NN), and the support vector machine (SVM).

In some pattern recognition tasks, the dimension of the sample space is larger than the number of samples in the training set. This is known as the "small sample size problem". Linear discriminant analysis (LDA) techniques cannot be applied directly to the small sample size case. The small sample size problem is also encountered when kernel approaches are used for recognition. In this paper, we attempt to answer the question of "How should one choose the optimal projection vectors for feature extraction in the small sample size case?" Based on our findings, we propose a new method called the kernel discriminative common vector method. In this method, we first nonlinearly map the original input space to an implicit higher dimensional feature space, in which the data are hoped to be linearly separable. Then, the optimal projection vectors are computed in this transformed space. The proposed method yields an optimal solution for maximizing a modified Fisher's linear discriminant criterion, discussed in the paper. Thus, under certain conditions, a 100% recognition rate is guaranteed for the training set samples. Experiments on test data also show that, in many situations, the generalization performance of the proposed method compares favorably with other kernel approaches.

In pattern recognition, feature extraction techniques are widely employed to reduce the dimensionality of data and to enhance the discriminatory information. Principal component analysis (PCA) and linear discriminant analysis (LDA) are the two most popular linear dimensionality reduction methods. However, PCA is not very effective for the extraction of the most discriminant features, and LDA is not stable due to the small sample size problem . In this paper, we propose some new (linear and nonlinear) feature extractors based on maximum margin criterion (MMC). Geometrically, feature extractors based on MMC maximize the (average) margin between classes after dimensionality reduction. It is shown that MMC can represent class separability better than PCA. As a connection to LDA, we may also derive LDA from MMC by incorporating some constraints. By using some other constraints, we establish a new linear feature extractor that does not suffer from the small sample size problem, which is known to cause serious stability problems for LDA. The kernelized (nonlinear) counterpart of this linear feature extractor is also established in the paper. Our extensive experiments demonstrate that the new feature extractors are effective, stable, and efficient.

A novel feature selection method using the concept of mutual information (MI) is proposed in this paper. In all MI based feature selection methods, effective and efficient estimation of high-dimensional MI is crucial. In this paper, a pruned Parzen window estimator and the quadratic mutual information (QMI) are combined to address this problem. The results show that the proposed approach can estimate the MI in an effective and efficient way. With this contribution, a novel feature selection method is developed to identify the salient features one by one. Also, the appropriate feature subsets for classification can be reliably estimated. The proposed methodology is thoroughly tested in four different classification applications in which the number of features ranged from less than 10 to over 15,000. The presented results are very promising and corroborate the contribution of the proposed feature selection methodology.

The decision-making process of many binary classification systems is based on the likelihood ratio (LR) scores of test patterns. This paper shows that LR scores can be expressed in terms of the similarity between the supervectors (SVs) formed by stacking the mean vectors of Gaussian mixture models corresponding to the test patterns, the target model, and the background model. By interpreting the support vector machine (SVM) kernels as a specific similarity (or discriminant) function between SVs, this paper shows that LR scoring is a special case of SVM scoring and that most sequence kernels can be obtained by assuming a specific form for the similarity function of SVs. This paper further shows that this assumption can be relaxed to derive a new general kernel. The kernel function is general in that it is a linear combination of any kernels belonging to the reproducing kernel Hilbert space. The combination weights are obtained by optimizing the ability of a discriminant function to separate the positive and negative classes using either regression analysis or SVM training. The idea was applied to both high-and low-level speaker verification. In both cases, results show that the proposed kernels achieve better performance than several state-of-the-art sequence kernels. Further performance enhancement was also observed when the high-level scores were combined with acoustic scores.

This paper proposes an uncorrelated multilinear principal component analysis (UMPCA) algorithm for unsupervised subspace learning of tensorial data. It should be viewed as a multilinear extension of the classical principal component analysis (PCA) framework. Through successive variance maximization, UMPCA seeks a tensor-to-vector projection (TVP) that captures most of the variation in the original tensorial input while producing uncorrelated features. The solution consists of sequential iterative steps based on the alternating projection method. In addition to deriving the UMPCA framework, this work offers a way to systematically determine the maximum number of uncorrelated multilinear features that can be extracted by the method. UMPCA is compared against the baseline PCA solution and its five state-of-the-art multilinear extensions, namely two-dimensional PCA (2DPCA), concurrent subspaces analysis (CSA), tensor rank-one decomposition (TROD), generalized PCA (GPCA), and multilinear PCA (MPCA), on the tasks of unsupervised face and gait recognition. Experimental results included in this paper suggest that UMPCA is particularly effective in determining the low-dimensional projection space needed in such recognition tasks.

An unsupervised competitive neural network for efficient clustering of Gaussian probability density function (GPDF) data of continuous density hidden Markov models (CDHMMs) is proposed in this paper. The proposed unsupervised competitive neural network, called the divergence-based centroid neural network (DCNN), employs the divergence measure as its distance measure and utilizes the statistical characteristics of observation densities in the HMM for speech recognition problems. While the conventional clustering algorithms used for the vector quantization (VQ) codebook design utilize only the mean values of the observation densities in the HMM, the proposed DCNN utilizes both the mean and the covariance values. When compared with other conventional unsupervised neural networks, the DCNN successfully allocates more code vectors to the regions where GPDF data are densely distributed while it allocates fewer code vectors to the regions where GPDF data are sparsely distributed. When applied to Korean monophone recognition problems as a tool to reduce the size of the codebook, the DCNN reduced the number of GPDFs used for code vectors by 65.3% while preserving recognition accuracy. Experimental results with a divergence-based k-means algorithm and a divergence-based self-organizing map algorithm are also presented in this paper for a performance comparison.

Earlier clustering techniques such as the modified learning vector quantization (MLVQ) and the fuzzy Kohonen partitioning (FKP) techniques have focused on the derivation of a certain set of parameters so as to define the fuzzy sets in terms of an algebraic function. The fuzzy membership functions thus generated are uniform, normal, and convex. Since any irregular training data is clustered into uniform fuzzy sets (Gaussian, triangular, or trapezoidal), the clustering may not be exact and some amount of information may be lost. In this paper, two clustering techniques using a Kohonen-like self-organizing neural network architecture, namely, the unsupervised discrete clustering technique (UDCT) and the supervised discrete clustering technique (SDCT), are proposed. The UDCT and SDCT algorithms reduce this data loss by introducing nonuniform, normal fuzzy sets that are not necessarily convex. The training data range is divided into discrete points at equal intervals, and the membership value corresponding to each discrete point is generated. Hence, the fuzzy sets obtained contain pairs of values, each pair corresponding to a discrete point and its membership grade. Thus, it can be argued that fuzzy membership functions generated using this kind of a discrete methodology provide a more accurate representation of the actual input data. This fact has been demonstrated by comparing the membership functions generated by the UDCT and SDCT algorithms against those generated by the MLVQ, FKP, and pseudofuzzy Kohonen partitioning (PFKP) algorithms. In addition to these clustering techniques, a novel pattern classifying network called the Yager fuzzy neural network (FNN) is proposed in this paper. This network corresponds completely to the Yager inference rule and exhibits remarkable generalization abilities. A modified version of the pseudo-outer product (POP)-Yager FNN called the modified Yager FNN is introduced that eliminates the drawbacks of the earlier network and yi- elds superior performance. Extensive experiments have been conducted to test the effectiveness of these two networks, using various clustering algorithms. It follows that the SDCT and UDCT clustering algorithms are particularly suited to networks based on the Yager inference rule.

A novel radial basis function neural network for discriminant analysis is presented in this paper. In contrast to many other researches, this work focuses on the exploitation of the weight structure of radial basis function neural networks using the Bayesian method. It is expected that the performance of a radial basis function neural network with a well-explored weight structure can be improved. As the weight structure of a radial basis function neural network is commonly unknown, the Bayesian method is, therefore, used in this paper to study this a priori structure. Two weight structures are investigated in this study, i.e., a single-Gaussian structure and a two-Gaussian structure. An expectation-maximization learning algorithm is used to estimate the weights. The simulation results showed that the proposed radial basis function neural network with a weight structure of two Gaussians outperformed the other algorithms.

Nonnegative matrix factorization (NMF) is an unsupervised learning method useful in various applications including image processing and semantic analysis of documents. This paper focuses on symmetric NMF (SNMF), which is a special case of NMF decomposition. Three parallel multiplicative update algorithms using level 3 basic linear algebra subprograms directly are developed for this problem. First, by minimizing the Euclidean distance, a multiplicative update algorithm is proposed, and its convergence under mild conditions is proved. Based on it, we further propose another two fast parallel methods: α-SNMF and β -SNMF algorithms. All of them are easy to implement. These algorithms are applied to probabilistic clustering. We demonstrate their effectiveness for facial image clustering, document categorization, and pattern clustering in gene expression.

The linear discriminant analysis (LDA) is a very popular linear feature extraction approach. The algorithms of LDA usually perform well under the following two assumptions. The first assumption is that the global data structure is consistent with the local data structure. The second assumption is that the input data classes are Gaussian distributions. However, in real-world applications, these assumptions are not always satisfied. In this paper, we propose an improved LDA framework, the local LDA (LLDA), which can perform well without needing to satisfy the above two assumptions. Our LLDA framework can effectively capture the local structure of samples. According to different types of local data structure, our LLDA framework incorporates several different forms of linear feature extraction approaches, such as the classical LDA and principal component analysis. The proposed framework includes two LLDA algorithms: a vector-based LLDA algorithm and a matrix-based LLDA (MLLDA) algorithm. MLLDA is directly applicable to image recognition, such as face recognition. Our algorithms need to train only a small portion of the whole training set before testing a sample. They are suitable for learning large-scale databases especially when the input data dimensions are very high and can achieve high classification accuracy. Extensive experiments show that the proposed algorithms can obtain good classification results.

This paper presents a new wrapper-based feature selection method for support vector regression (SVR) using its probabilistic predictions. The method computes the importance of a feature by aggregating the difference, over the feature space, of the conditional density functions of the SVR prediction with and without the feature. As the exact computation of this importance measure is expensive, two approximations are proposed. The effectiveness of the measure using these approximations, in comparison to several other existing feature selection methods for SVR, is evaluated on both artificial and real-world problems. The result of the experiments show that the proposed method generally performs better than, or at least as well as, the existing methods, with notable advantage when the dataset is sparse.

This paper proposes a hybrid multilogistic methodology, named logistic regression using initial and radial basis function (RBF) covariates. The process for obtaining the coefficients is carried out in three steps. First, an evolutionary programming (EP) algorithm is applied, in order to produce an RBF neural network (RBFNN) with a reduced number of RBF transformations and the simplest structure possible. Then, the initial attribute space (or, as commonly known as in logistic regression literature, the covariate space) is transformed by adding the nonlinear transformations of the input variables given by the RBFs of the best individual in the final generation. Finally, a maximum likelihood optimization method determines the coefficients associated with a multilogistic regression model built in this augmented covariate space. In this final step, two different multilogistic regression algorithms are applied: one considers all initial and RBF covariates (multilogistic initial-RBF regression) and the other one incrementally constructs the model and applies cross validation, resulting in an automatic covariate selection [simplelogistic initial-RBF regression (SLIRBF)]. Both methods include a regularization parameter, which has been also optimized. The methodology proposed is tested using 18 benchmark classification problems from well-known machine learning problems and two real agronomical problems. The results are compared with the corresponding multilogistic regression methods applied to the initial covariate space, to the RBFNNs obtained by the EP algorithm, and to other probabilistic classifiers, including different RBFNN design methods [e.g., relaxed variable kernel density estimation, support vector machines, a sparse classifier (sparse multinomial logistic regression)] and a procedure similar to SLIRBF but using product unit basis functions. The SLIRBF models are found to be competitive when compared with the corresponding multilogistic regression methods and the RBFEP method. A measure of statistical significance is used, which indicates that SLIRBF reaches the state of the art.

In this paper, we consider the problem of constructing accurate and flexible statistical representations for count data, which we often confront in many areas such as data mining, computer vision, and information retrieval. In particular, we analyze and compare several generative approaches widely used for count data clustering, namely multinomial, multinomial Dirichlet, and multinomial generalized Dirichlet mixture models. Moreover, we propose a clustering approach via a mixture model based on a composition of the Liouville family of distributions, from which we select the Beta-Liouville distribution, and the multinomial. The novel proposed model, which we call multinomial Beta-Liouville mixture, is optimized by deterministic annealing expectation-maximization and minimum description length, and strives to achieve a high accuracy of count data clustering and model selection. An important feature of the multinomial Beta-Liouville mixture is that it has fewer parameters than the recently proposed multinomial generalized Dirichlet mixture. The performance evaluation is conducted through a set of extensive empirical experiments, which concern text and image texture modeling and classification and shape modeling, and highlights the merits of the proposed models and approaches.

This paper discusses the problem of what kind of learning model is suitable for the tasks of feature extraction for data representation and suggests two evaluation criteria for nonlinear feature extractors: reconstruction error minimization and similarity preservation. Based on the suggested evaluation criteria, a new type of principal curve-similarity preserving principal curve (SPPC) is proposed. SPPCs minimize the reconstruction error under the condition that the similarity between similar samples are preserved in the extracted features, thus giving researchers effective and reliable cognition of the inner structure of data sets. The existence and properties of SPPCs are analyzed; a practical learning algorithm is proposed and high dimensional extensions of SPPCs are also discussed. Experimental results show the virtues of SPPCs in preserving inner structures of data sets and discovering manifolds with high nonlinearity.

The weight-decay technique is an effective approach to handle overfitting and weight fault. For fault-free networks, without an appropriate value of decay parameter, the trained network is either overfitted or underfitted. However, many existing results on the selection of decay parameter focus on fault-free networks only. It is well known that the weight-decay method can also suppress the effect of weight fault. For the faulty case, using a test set to select the decay parameter is not practice because there are huge number of possible faulty networks for a trained network. This paper develops two mean prediction error (MPE) formulae for predicting the performance of faulty radial basis function (RBF) networks. Two fault models, multiplicative weight noise and open weight fault, are considered. Our MPE formulae involve the training error and trained weights only. Besides, in our method, we do not need to generate a huge number of faulty networks to measure the test error for the fault situation. The MPE formulae allow us to select appropriate values of decay parameter for faulty networks. Our experiments showed that, although there are small differences between the true test errors (from the test set) and the MPE values, the MPE formulae can accurately locate the appropriate value of the decay parameter for minimizing the true test error of faulty networks.

Recently, the minimum error entropy (MEE) criterion has been used as an information theoretic alternative to traditional mean-square error criterion in supervised learning systems. MEE yields nonquadratic, nonconvex performance surface even for adaptive linear neuron (ADALINE) training, which complicates the theoretical analysis of the method. In this paper, we develop a unified approach for mean-square convergence analysis for ADALINE training under MEE criterion. The weight update equation is formulated in the form of block-data. Based on a block version of energy conservation relation, and under several assumptions, we carry out the mean-square convergence analysis of this class of adaptation algorithm, including mean-square stability, mean-square evolution (transient behavior) and the mean-square steady-state performance. Simulation experimental results agree with the theoretical predictions very well.

The finite mixture model is widely used in various statistical learning problems. However, the model obtained may contain a large number of components, making it inefficient in practical applications. In this paper, we propose to simplify the mixture model by minimizing an upper bound of the approximation error between the original and the simplified model, under the use of the L (2) distance measure. This is achieved by first grouping similar components together and then performing local fitting through function approximation. The simplified model obtained can then be used as a replacement of the original model to speed up various algorithms involving mixture models during training (e.g., Bayesian filtering, belief propagation) and testing [e.g., kernel density estimation, support vector machine (SVM) testing]. Encouraging results are observed in the experiments on density estimation, clustering-based image segmentation, and simplification of SVM decision functions.

This paper develops a probabilistic approach to recursive second-order training of recurrent neural networks (RNNs) for improved time-series modeling. A general recursive Bayesian Levenberg-Marquardt algorithm is derived to sequentially update the weights and the covariance (Hessian) matrix. The main strengths of the approach are a principled handling of the regularization hyperparameters that leads to better generalization, and stable numerical performance. The framework involves the adaptation of a noise hyperparameter and local weight prior hyperparameters, which represent the noise in the data and the uncertainties in the model parameters. Experimental investigations using artificial and real-world data sets show that RNNs equipped with the proposed approach outperform standard real-time recurrent learning and extended Kalman training algorithms for recurrent networks, as well as other contemporary nonlinear neural models, on time-series modeling.

In this paper, we propose a clustering algorithm based on both Dirichlet processes and generalized Dirichlet distribution which has been shown to be very flexible for proportional data modeling. Our approach can be viewed as an extension of the finite generalized Dirichlet mixture model to the infinite case. The extension is based on nonparametric Bayesian analysis. This clustering algorithm does not require the specification of the number of mixture components to be given in advance and estimates it in a principled manner. Our approach is Bayesian and relies on the estimation of the posterior distribution of clusterings using Gibbs sampler. Through some applications involving real-data classification and image databases categorization using visual words, we show that clustering via infinite mixture models offers a more powerful and robust performance than classic finite mixtures.

In this paper, we address the problem of estimating the parameters of Gaussian mixture models. Although the expectation-maximization (EM) algorithm yields the maximum-likelihood (ML) solution, its sensitivity to the selection of the starting parameters is well-known and it may converge to the boundary of the parameter space. Furthermore, the resulting mixture depends on the number of selected components, but the optimal number of kernels may be unknown beforehand. We introduce the use of the entropy of the probability density function (pdf) associated to each kernel to measure the quality of a given mixture model with a fixed number of kernels. We propose two methods to approximate the entropy of each kernel and a modification of the classical EM algorithm in order to find the optimum number of components of the mixture. Moreover, we use two stopping criteria: a novel global mixture entropy-based criterion called Gaussianity deficiency (GD) and a minimum description length (MDL) principle-based one. Our algorithm, called entropy-based EM (EBEM), starts with a unique kernel and performs only splitting by selecting the worst kernel attending to GD. We have successfully tested it in probability density estimation, pattern classification, and color image segmentation. Experimental results improve the ones of other state-of-the-art model order selection methods.

This paper proposes an uncorrelated multilinear discriminant analysis (UMLDA) framework for the recognition of multidimensional objects, known as tensor objects. Uncorrelated features are desirable in recognition tasks since they contain minimum redundancy and ensure independence of features. The UMLDA aims to extract uncorrelated discriminative features directly from tensorial data through solving a tensor-to-vector projection. The solution consists of sequential iterative processes based on the alternating projection method, and an adaptive regularization procedure is incorporated to enhance the performance in the small sample size (SSS) scenario. A simple nearest-neighbor classifier is employed for classification. Furthermore, exploiting the complementary information from differently initialized and regularized UMLDA recognizers, an aggregation scheme is adopted to combine them at the matching score level, resulting in enhanced generalization performance while alleviating the regularization parameter selection problem. The UMLDA-based recognition algorithm is then empirically shown on face and gait recognition tasks to outperform four multilinear subspace solutions (MPCA, DATER, GTDA, TR1DA) and four linear subspace solutions (Bayesian, LDA, ULDA, R-JD-LDA).

Model selection in kernel linear discriminant analysis (KLDA) refers to the selection of appropriate parameters of a kernel function and the regularizer. By following the principle of maximum information preservation , this paper formulates the model selection problem as a problem of selecting an optimal kernel-induced space in which different classes are maximally separated from each other. A scatter-matrix-based criterion is developed to measure the “goodness” of a kernel-induced space, and the kernel parameters are tuned by maximizing this criterion. This criterion is computationally efficient and is differentiable with respect to the kernel parameters. Compared with the leave-one-out (LOO) or $k$ -fold cross validation (CV), the proposed approach can achieve a faster model selection, especially when the number of training samples is large or when many kernel parameters need to be tuned. To tune the regularization parameter in the KLDA, our criterion is used together with the method proposed by Saadi (2004). Experiments on benchmark data sets verify the effectiveness of this model selection approach.

The insufficiency of using only second-order statistics and premise of exploiting higher order statistics of the data has been well understood, and more advanced objectives including higher order statistics, especially those stemming from information theory, such as error entropy minimization, are now being studied and applied in many contexts of machine learning and signal processing. In the adaptive system training context, the main drawback of utilizing output error entropy as compared to correlation-estimation-based second-order statistics is the computational load of the entropy estimation, which is usually obtained via a plug-in kernel estimator. Sample-spacing estimates offer computationally inexpensive entropy estimators; however, resulting estimates are not differentiable, hence, not suitable for gradient-based adaptation. In this brief paper, we propose a nonparametric entropy estimator that captures the desirable properties of both approaches. The resulting estimator yields continuously differentiable estimates with a computational complexity at the order of those of the sample-spacing techniques. The proposed estimator is compared with the kernel density estimation (KDE)-based entropy estimator in the supervised neural network training framework with computation time and performance comparisons.

In this paper, a new kernel-based method for data visualization and dimensionality reduction is proposed. A reference point is considered corresponding to additional constraints taken in the problem formulation. In contrast with the class of kernel eigenmap methods, the solution (coordinates in the low-dimensional space) is characterized by a linear system instead of an eigenvalue problem. The kernel maps with a reference point are generated from a least squares support vector machine (LS-SVM) core part that is extended with an additional regularization term for preserving local mutual distances together with reference point constraints. The kernel maps possess primal and dual model representations and provide out-of-sample extensions, e.g., for validation-based tuning. The method is illustrated on toy problems and real-life data sets.

In this paper, we present a neural classifier algorithm that locally approximates the decision surface of labeled data by a patchwork of separating hyperplanes, which are arranged under certain topological constraints similar to those of self-organizing maps (SOMs). We take advantage of the fact that these boundaries can often be represented by linear ones connected by a low-dimensional nonlinear manifold, thus influencing the placement of the separators. The resulting classifier allows for a voting scheme that averages over the classification results of neighboring hyper- planes. Our algorithm is computationally efficient both in terms of training and classification. Further, we present a model selection method to estimate the topology of the classification boundary. We demonstrate the algorithm's usefulness on several artificial and real-world data sets and compare it to the state-of-the-art supervised learning algorithms.

Hard-margin support vector machines (HM-SVMs) suffer from getting overfitting in the presence of noise. Soft-margin SVMs deal with this problem by introducing a regularization term and obtain a state-of-the-art performance. However, this disposal leads to a relatively high computational cost. In this paper, an alternative method, greedy stagewise algorithm for SVMs, named GS-SVMs, is presented to cope with the overfitting of HM-SVMs without employing the regularization term. The most attractive property of GS-SVMs is that its computational complexity in the worst case only scales quadratically with the size of training samples. Experiments on the large data sets with up to 400,000 training samples demonstrate that GS-SVMs can be faster than LIBSVM 2.83 without sacrificing the accuracy. Finally, we employ statistical learning theory to analyze the empirical results, which shows that the success of GS-SVMs lies in that its early stopping rule can act as an implicit regularization term.

Plenty of methods have been proposed in order to discover latent variables (features) in data sets. Such approaches include the principal component analysis (PCA), independent component analysis (ICA), factor analysis (FA), etc., to mention only a few. A recently investigated approach to decompose a data set with a given dimensionality into a lower dimensional space is the so-called nonnegative matrix factorization (NMF). Its only requirement is that both decomposition factors are nonnegative. To approximate the original data, the minimization of the NMF objective function is performed in the Euclidean space, where the difference between the original data and the factors can be minimized by employing L(2)-norm. In this paper, we propose a generalization of the NMF algorithm by translating the objective function into a Hilbert space (also called feature space) under nonnegativity constraints. With the help of kernel functions, we developed an approach that allows high-order dependencies between the basis images while keeping the nonnegativity constraints on both basis images and coefficients. Two practical applications, namely, facial expression and face recognition, show the potential of the proposed approach.

The adaptive-subspace self-organizing map (ASSOM) is useful for invariant feature generation and visualization. However, the learning procedure of the ASSOM is slow. In this paper, two fast implementations of the ASSOM are proposed to boost ASSOM learning based on insightful discussions of the basis rotation operator of ASSOM. We investigate the objective function approximately maximized by the classical rotation operator. We then explore a sequence of two schemes to apply the proposed ASSOM implementations to saliency-based invariant feature construction for image classification. In the first scheme, a cumulative activity map computed from a single ASSOM is used as descriptor of the input image. In the second scheme, we use one ASSOM for each image category and a joint cumulative activity map is calculated as the descriptor. Both schemes are evaluated on a subset of the Corel photo database with ten classes. The multi-ASSOM scheme is favored. It is also applied to adult image filtering and shows promising results.

In this paper, we propose a powerful symmetric radial basis function (RBF) classifier for nonlinear detection in the so-called "overloaded" multiple-antenna-aided communication systems. By exploiting the inherent symmetry property of the optimal Bayesian detector, the proposed symmetric RBF classifier is capable of approaching the optimal classification performance using noisy training data. The classifier construction process is robust to the choice of the RBF width and is computationally efficient. The proposed solution is capable of providing a signal-to-noise ratio (SNR) gain in excess of 8 dB against the powerful linear minimum bit error rate (BER) benchmark, when supporting four users with the aid of two receive antennas or seven users with four receive antenna elements.

In this paper, we propose a novel large margin classifier, called the maxi-min margin machine M(4). This model learns the decision boundary both locally and globally. In comparison, other large margin classifiers construct separating hyperplanes only either locally or globally. For example, a state-of-the-art large margin classifier, the support vector machine (SVM), considers data only locally, while another significant model, the minimax probability machine (MPM), focuses on building the decision hyperplane exclusively based on the global information. As a major contribution, we show that SVM yields the same solution as M(4) when data satisfy certain conditions, and MPM can be regarded as a relaxation model of M(4). Moreover, based on our proposed local and global view of data, another popular model, the linear discriminant analysis, can easily be interpreted and extended as well. We describe the M(4) model definition, provide a geometrical interpretation, present theoretical justifications, and propose a practical sequential conic programming method to solve the optimization problem. We also show how to exploit Mercer kernels to extend M(4) for nonlinear classifications. Furthermore, we perform a series of evaluations on both synthetic data sets and real-world benchmark data sets. Comparison with SVM and MPM demonstrates the advantages of our new model.

In this paper, we modify the fuzzy ARTMAP (FA) neural network (NN) using the Bayesian framework in order to improve its classification accuracy while simultaneously reduce its category proliferation. The proposed algorithm, called Bayesian ARTMAP (BA), preserves the FA advantages and also enhances its performance by the following: (1) representing a category using a multidimensional Gaussian distribution, (2) allowing a category to grow or shrink, (3) limiting a category hypervolume, (4) using Bayes' decision theory for learning and inference, and (5) employing the probabilistic association between every category and a class in order to predict the class. In addition, the BA estimates the class posterior probability and thereby enables the introduction of loss and classification according to the minimum expected loss. Based on these characteristics and using synthetic and 20 real-world databases, we show that the BA outperformes the FA, either trained for one epoch or until completion, with respect to classification accuracy, sensitivity to statistical overlapping, learning curves, expected loss, and category proliferation.

This paper is concerned with the problem of local and global asymptotic stability for a class of discrete-time recurrent neural networks, which provide discrete-time analogs to their continuous-time counterparts, i.e., continuous-time recurrent neural networks with distributed delay. Some stability criteria, which include some existing results as their special cases, are derived. A discussion about the dynamical consistence of discrete-time neural networks versus their continuous-time counterparts is provided. An unconventional finite difference method is proposed and an example is also given to show the effectiveness of the method.

Independent component analysis (ICA) is currently the most popularly used approach to blind source separation (BSS), the problem of recovering unknown source signals when their mixtures are observed but the actual mixing process is unknown. Many ICA algorithms assume that a fixed set of source signals consistently exists in mixtures throughout the time-series to be examined. However, real-world signals often have such difficult nonstationarity that each source signal abruptly appears or disappears, thus the set of active sources dynamically changes with time. In this paper, we propose switching ICA (SwICA), which focuses on such situations. The proposed approach is based on the noisy ICA formulated as a generative model. We employ a special type of hidden Markov model (HMM) to represent such prior knowledge that the source may abruptly appear or disappear with time. The special HMM setting t hen provides an effect ofvariable selection in a dynamic way. We use the variational Bayes (VB) method to derive an effective approximation of Bayesian inference for this model. In simulation experiments using artificial and realistic source signals, the proposed method exhibited performance superior to existing methods, especially in the presence of noise. The compared methods include the natural-gradient ICA with a nonholonomic constraint, and the existing ICA method incorporating an HMM source model, which aims to deal with general nonstationarities that may exist in source signals. In addition, the proposed method could successfully recover the source signals even when the total number of true sources was overestimated or was larger than that of mixtures. We also propose a modification of the basic Markov model into a semi-Markov model, and show that the semi-Markov one is more effective for robust estimation of the source appearance.

International audience; Mapping high-dimensional data in a low-dimensional space, for example, for visualization, is a problem of increasingly major concern in data analysis. This paper presents data-driven high-dimensional scaling (DD-HDS), a nonlinear mapping method that follows the line of multidimensional scaling (MDS) approach, based on the preservation of distances between pairs of data. It improves the performance of existing competitors with respect to the representation of high-dimensional data, in two ways. It introduces (1) a specific weighting of distances between data taking into account the concentration of measure phenomenon and (2) a symmetric handling of short distances in the original and output spaces, avoiding false neighbor representations while still allowing some necessary tears in the original distribution. More precisely, the weighting is set according to the effective distribution of distances in the data set, with the exception of a single user-defined parameter setting the tradeoff between local neighborhood preservation and global mapping. The optimization of the stress criterion designed for the mapping is realized by "force-directed placement" (FDP). The mappings of low- and high-dimensional data sets are presented as illustrations of the features and advantages of the proposed algorithm. The weighting function specific to high-dimensional data and the symmetric handling of short distances can be easily incorporated in most distance preservation-based nonlinear dimensionality reduction methods.

This paper proposes polytope ARTMAP (PTAM), an adaptive resonance theory (ART) network for classification tasks which does not use the vigilance parameter. This feature is due to the geometry of categories in PTAM, which are irregular polytopes whose borders approximate the borders among the output predictions. During training, the categories expand only towards the input pattern without category overlap. The category expansion in PTAM is naturally limited by the other categories, and not by the category size, so the vigilance is not necessary. PTAM works in a fully automatic way for pattern classification tasks, without any parameter tuning, so it is easier to employ for nonexpert users than other classifiers. PTAM achieves lower error than the leading ART networks on a complete collection of benchmark data sets, except for noisy data, without any parameter optimization.

A Bayesian Gaussian process (GP) modeling approach has recently been introduced to model-based control strategies. The estimate of the variance of the predicted output is the most useful advantage of GPs in comparison to neural networks (NNs) and fuzzy models. However, the GP model is computationally demanding and nontransparent. To reduce the computation load and increase transparency, a local linear GP model network is proposed in this paper. The proposed methodology combines the local model network principle with the GP prior approach. A novel algorithm for structure determination and optimization is introduced, which is widely applicable to the training of local model networks. The modeling procedure of the local linear GP (LGP) model network is demonstrated on an example of a nonlinear laboratory scale process rig.

In this paper, we present a novel scheme for linear feature extraction in classification. The method is based on the maximization of the mutual information (MI) between the features extracted and the classes. The sum of the MI corresponding to each of the features is taken as an heuristic that approximates the MI of the whole output vector. Then, a component-by-component gradient-ascent method is proposed for the maximization of the MI, similar to the gradient-based entropy optimization used in independent component analysis (ICA). The simulation results show that not only is the method competitive when compared to existing supervised feature extraction methods in all cases studied, but it also remarkably outperform them when the data are characterized by strongly nonlinear boundaries between classes.

In this paper, we present two fast sparse approximation schemes for least squares support vector machine (LS-SVM), named FSALS-SVM and PFSALS-SVM, to overcome the limitation of LS-SVM that it is not applicable to large data sets and to improve test speed. FSALS-SVM iteratively builds the decision function by adding one basis function from a kernel-based dictionary at one time. The process is terminated by using a flexible and stable epsilon insensitive stopping criterion. A probabilistic speedup scheme is employed to further improve the speed of FSALS-SVM and the resulting classifier is named PFSALS-SVM. Our algorithms are of two compelling features: low complexity and sparse solution. Experiments on benchmark data sets show that our algorithms obtain sparse classifiers at a rather low cost without sacrificing the generalization performance.

In this paper, a novel exemplar-based constructive approach using kernels is proposed for simultaneous pattern classification and multidomain pattern association tasks. The kernel networks are constructed on a modular basis by a simple one-shot self-structuring algorithm motivated from the traditional Hebbian principle and then, they act as the flexible memory capable of generalization for the respective classes. In the self-structuring kernel memory (SSKM), any arduous and iterative network parameter tuning is not involved for establishing the weight connections during the construction, unlike conventional approaches, and thereby, it is considered that the networks do not inherently suffer from the associated numerical instability. Then, the approach is extended for multidomain pattern association, in which a particular domain input cannot only activate some kernel units (KUs) but also the kernels in other domain(s) via the cross-domain connection(s) in between. Thereby, the SSKM can be regarded as a simultaneous pattern classifier and associator. In the simulation study for pattern classification, it is justified that an SSKM consisting of distinct kernel networks can yield relatively compact-sized pattern classifiers, while preserving a reasonably high generalization capability, in comparison with the approach using support vector machines (SVMs).

While distance function learning for supervised learning tasks has a long history, extending it to learning tasks with weaker supervisory information has only been studied recently. In particular, some methods have been proposed for semisupervised metric learning based on pairwise similarity or dissimilarity information. In this paper, we propose a kernel approach for semisupervised metric learning and present in detail two special cases of this kernel approach. The metric learning problem is thus formulated as an optimization problem for kernel learning. An attractive property of the optimization problem is that it is convex and, hence, has no local optima. While a closed-form solution exists for the first special case, the second case is solved using an iterative majorization procedure to estimate the optimal solution asymptotically. Experimental results based on both synthetic and real-world data show that this new kernel approach is promising for nonlinear metric learning.

This paper presents a new classifier called total margin-based adaptive fuzzy support vector machines (TAF-SVM) that deals with several problems that may occur in support vector machines (SVMs) when applied to the face recognition. The proposed TAF-SVM not only solves the overfitting problem resulted from the outlier with the approach of fuzzification of the penalty, but also corrects the skew of the optimal separating hyperplane due to the very imbalanced data sets by using different cost algorithm. In addition, by introducing the total margin algorithm to replace the conventional soft margin algorithm, a lower generalization error bound can be obtained. Those three functions are embodied into the traditional SVM so that the TAF-SVM is proposed and reformulated in both linear and nonlinear cases. By using two databases, the Chung Yuan Christian University (CYCU) multiview and the facial recognition technology (FERET) face databases, and using the kernel Fisher's discriminant analysis (KFDA) algorithm to extract discriminating face features, experimental results show that the proposed TAF-SVM is superior to SVM in terms of the face-recognition accuracy. The results also indicate that the proposed TAF-SVM can achieve smaller error variances than SVM over a number of tests such that better recognition stability can be obtained.