Non-negative matrix factorization (NMF) is a recently popularized technique for learning partsbased, linear representations of non-negative data. The traditional NMF is optimized under the Gaussian noise or Poisson noise assumption, and hence not suitable if the data are grossly corrupted. To improve the robustness of NMF, a novel algorithm named robust nonnegative matrix factorization (RNMF) is proposed in this paper. We assume that some entries of the data matrix may be arbitrarily corrupted, but the corruption is sparse. RNMF decomposes the non-negative data matrix as the summation of one sparse error matrix and the product of two non-negative matrices. An efficient iterative approach is developed to solve the optimization problem of RNMF. We present experimental results on two face databases to verify the effectiveness of the proposed method.
Cerebellar model articulation controller (CMAC) is a popular associative memory neural network that imitates human’s cerebellum, which allows it to learn fast and carry out local generalization efficiently. This research aims to integrate evolutionary computation into fuzzy CMAC Bayesian Ying-Yang (FCMAC-BYY) learning, which is referred to as FCMAC-EBYY, to achieve a synergetic development in the search for optimal fuzzy sets and connection weights. Traditional evolutionary approaches are limited to small populations of short binary string length and as such are not suitable for neural network training, which involves a large searching space due to complex connections as well as real values. The methodology employed by FCMAC-EBYY is coevolution, in which a complex solution is decomposed into some pieces to be optimized in different populations/species and then assembled. The developed FCMAC-EBYY is compared with various neuro-fuzzy systems using a real application of traffic flow prediction.
Three Bayesian related approaches, namely, variational Bayesian (VB), minimum message length (MML) and Bayesian Ying-Yang (BYY) harmony learning, have been applied to automatically determining an appropriate number of components during learning Gaussian mixture model (GMM). This paper aims to provide a comparative investigation on these approaches with not only a Jeffreys prior but also a conjugate Dirichlet-Normal-Wishart (DNW) prior on GMM. In addition to adopting the existing algorithms either directly or with some modifications, the algorithm for VB with Jeffreys prior and the algorithm for BYY with DNW prior are developed in this paper to fill the missing gap. The performances of automatic model selection are evaluated through extensive experiments, with several empirical findings: 1) Considering priors merely on the mixing weights, each of three approaches makes biased mistakes, while considering priors on all the parameters of GMM makes each approach reduce its bias and also improve its performance. 2) As Jeffreys prior is replaced by the DNW prior, all the three approaches improve their performances. Moreover, Jeffreys prior makes MML slightly better than VB, while the DNW prior makes VB better than MML. 3) As the hyperparameters of DNW prior are further optimized by each of its own learning principle, BYY improves its performances while VB and MML deteriorate their performances when there are too many free hyper-parameters. Actually, VB and MML lack a good guide for optimizing the hyper-parameters of DNW prior. 4) BYY considerably outperforms both VB and MML for any type of priors and whether hyper-parameters are optimized. Being different from VB and MML that rely on appropriate priors to perform model selection, BYY does not highly depend on the type of priors. It has model selection ability even without priors and performs already very well with Jeffreys prior, and incrementally improves as Jeffreys prior is replaced by the DNW prior. Finally, all algorithms are applied on the Berkeley segmentation database of real world images. Again, BYY considerably outperforms both VB and MML, especially in detecting the objects of interest from a confusing background.
Based on the problem of detecting the number of signals, this paper provides a systematic empiricalinvestigation on model selection performances of several classical criteria and recently developed methods (including Akaike’s information criterion (AIC), Schwarz’s Bayesian information criterion, Bozdogan’s consistent AIC, Hannan-Quinn information criterion, Minka’s (MK) principal component analysis (PCA) criterion, Kritchman & Nadler’s hypothesis tests (KN), Perry & Wolfe’s minimax rank estimation thresholding algorithm (MM), and Bayesian Ying-Yang (BYY) harmony learning), by varying signal-to-noise ratio (SNR) and training sample size
With the rapid development of next generation deep sequencing technologies, sequencing cDNA reverse-transcribed from RNA molecules (RNA-Seq) has become a key approach in studying gene expression and transcriptomes. Because RNA-Seq does not rely on annotation of known genes, it provides the opportunity of discovering transcripts that have not been annotated in current databases. Studying the distribution of RNA-Seq signals and a systematic view on the potential new transcripts revealed from the signals is an important step toward the understanding of transcriptomes.
This paper presents a new discriminative approach for training Gaussian mixture models (GMMs) of hidden Markov models (HMMs) based acoustic model in a large vocabulary continuous speech recognition (LVCSR) system. This approach is featured by embedding a rival penalized competitive learning (RPCL) mechanism on the level of hidden Markov states. For every input, the correct identity state, called winner and obtained by the Viterbi force alignment, is enhanced to describe this input while its most competitive rival is penalized by de-learning, which makes GMMs-based states become more discriminative.Without the extensive computing burden required by typical discriminative learning methods for one-pass recognition of the training set, the new approach saves computing costs considerably. Experiments show that the proposed method has a good convergence with better performances than the classical maximum likelihood estimation (MLE) based method. Comparing with two conventional discriminative methods, the proposed method demonstrates improved generalization ability, especially when the test set is not well matched with the training set.
With the ever increasing complexity of industrial systems, model-based control has encountered difficulties and is facing problems, while the interest in data-based control has been booming. This paper gives an overview of data-based control, which divides it into two subfields, intelligent modeling and direct controller design. In the two subfields, some important methods concerning data-based control are intensively investigated. Within the framework of data-based modeling, main modeling technologies and control strategies are discussed, and then fundamental concepts and various algorithms are presented for the design of a data-based controller. Finally, some remaining challenges are suggested.
Radar high-resolution range profiles (HRRPs) are typical high-dimensional and interdimension dependently distributed data, the statistical modeling of which is a challenging task for HRRP-based target recognition. Supposing that HRRP samples are independent and jointly Gaussian distributed, a recent work [Du L, Liu H W, Bao Z. IEEE Transactions on Signal Processing, 2008, 56(5): 1931-1944] applied factor analysis (FA) to model HRRP data with a two-phase approach for model selection, which achieved satisfactory recognition performance. The theoretical analysis and experimental results reveal that there exists high temporal correlation among adjacent HRRPs. This paper is thus motivated to model the spatial and temporal structure of HRRP data simultaneously by employing temporal factor analysis (TFA) model. For a limited size of high-dimensional HRRP data, the two-phase approach for parameter learning and model selection suffers from intensive computation burden and deteriorated evaluation. To tackle these problems, this work adopts the Bayesian Ying-Yang (BYY) harmony learning that has automatic model selection ability during parameter learning. Experimental results show stepwise improved recognition and rejection performances from the two-phase learning based FA, to the two-phase learning based TFA and to the BYY harmony learning based TFA with automatic model selection. In addition, adding many extra free parameters to the classic FA model and thus becoming even worse in identifiability, the model of a general linear dynamical system is even inferior to the classic FA model.
Fine and sparse details appear in many quality inspection applications requiring machine vision. Especially on flat surfaces, such as paper or board, the details can be made detectable by oblique illumination. In this study, a general definition of such details is given by defining sufficient statistical properties from histograms. The statistical model allows simulation of data and comparison of methods designed for detail detection. Based on the definition, utilization of the existing thresholding methods is shown to be well motivated. The comparison shows that minimum error thresholding outperforms the other standard methods. Finally, the results are successfully applied to a paper printability inspection application, and the IGT picking assessment, in which small surface defects must be detected. The provided method and measurement system prototype provide automated assessment with results comparable to manual expert evaluations in this laborious task.
Classification of different weather conditions provides a first step support for outdoor scene modeling, which is a core component in many different applications of outdoor video analysis and computer vision. Features derived from intrinsic properties of the visual effects of different weather conditions contribute to successful classification. In this paper, features representing both the autocorrelation of pixel-wise intensities over time and the max directional length of rain streaks or snowflakes are proposed. Based on the autocorrelation of each pixel’s intensities over time, two temporal features are used for coarse classification of weather conditions according to their visual effects. On the other hand, features are extracted for fine classification of video clips with rain and snow. The classification results on 249 video clips associated with different weather conditions indicate the effectiveness of the extracted features, by using C-SVM as the classifier.
A 3D face recognition approach which uses principal axes registration (PAR) and three face representation features from the re-sampling depth image: Eigenfaces, Fisherfaces and Zernike moments is presented. The approach addresses the issue of 3D face registration instantly achieved by PAR. Because each facial feature has its own advantages, limitations and scope of use, different features will complement each other. Thus the fusing features can learn more expressive characterizations than a single feature. The support vector machine (SVM) is applied for classification. In this method, based on the complementarity between different features, weighted decision-level fusion makes the recognition system have certain fault tolerance. Experimental results show that the proposed approach achieves superior performance with the rank-1 recognition rate of 98.36% for GavabDB database.
Several meta-analyses were recently conducted in attempts to identify the core brain regions exhibiting pathological changes in schizophrenia, which could potentially act as disease markers. Based on the findings of these meta-analyses, we developed a multivariate pattern analysis method to classify schizophrenic patients and healthy controls using structural magnetic resonance imaging (sMRI) data. Independent component analysis (ICA) was used to decompose gray matter density images into a set of spatially independent components. Spatial multiple regression of a region of interest (ROI) mask with each of the components was then performed to determine pathological patterns, in which the voxels were taken as features for classification. After dimensionality reduction using principal component analysis (PCA), a nonlinear support vector machine (SVM) classifier was trained to discriminate schizophrenic patients from healthy controls. The performance of the classifier was tested using a 10-fold cross-validation strategy. Experimental results showed that two distinct spatial patterns displayed discriminative power for schizophrenia, which mainly included the prefrontal cortex (PFC) and subcortical regions respectively. It was found that simultaneous usage of these two patterns improved the classification performance compared to using either of them alone. Moreover, the two pathological patterns constitute a prefronto-subcortical network, suggesting that schizophrenia involves abnormalities in networks of brain regions.
Kernel-based clustering is supposed to provide a better analysis tool for pattern classification, which implicitly maps input samples to a high-dimensional space for improving pattern separability. For this implicit space map, the kernel trick is believed to elegantly tackle the problem of “curse of dimensionality”, which has actually been more challenging for kernel-based clustering in terms of computational complexity and classification accuracy, which traditional kernelized algorithms cannot effectively deal with. In this paper, we propose a novel kernel clustering algorithm, called KFCM-III, for this problem by replacing the traditional isotropic Gaussian kernel with the anisotropic kernel formulated by Mahalanobis distance. Moreover, a reduced-set represented kernelized center has been employed for reducing the computational complexity of KFCM-I algorithm and circumventing the model deficiency of KFCM-II algorithm. The proposed KFCMIII has been evaluated for segmenting magnetic resonance imaging (MRI) images. For this task, an image intensity inhomogeneity correction is employed during image segmentation process. With a scheme called preclassification, the proposed intensity correction scheme could further speed up image segmentation. The experimental results on public image data show the superiorities of KFCM-III.
The aims of this study are to develop the color density concept and to propose the color density based color difference formulas. The color density is defined using the metric coefficients that are based on the discrimination ellipses and the locations of the colors in the color space. The ellipse sets are the MacAdam ellipses in the CIE 1931
The design of the infrastructure for Chinese Web (CWI), a prototype system aimed at forum data analysis, is introduced. CWI takes a best effort approach. 1) It tries its best to extract or annotate semantics over the web data. 2) It provides flexible schemes for users to transform the web data into eXtensible Markup Language (XML) forms with more semantic annotations that are more friendly for further analytical tasks. 3) A distributed graph repository, called DISGR is used as backend for management of web data. The paper introduces the design issues, reports the progress of the implementation, and discusses the research issues that are under study.