By Topic

Machine Learning for Signal Processing (MLSP), 2011 IEEE International Workshop on

Date 18-21 Sept. 2011

Filter Results

Displaying Results 1 - 25 of 102
  • [Title page]

    Page(s): 1
    Save to Project icon | Request Permissions | PDF file iconPDF (154 KB)  
    Freely Available from IEEE
  • [Copyright notice]

    Page(s): 1
    Save to Project icon | Request Permissions | PDF file iconPDF (192 KB)  
    Freely Available from IEEE
  • Preface

    Page(s): 1
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (345 KB) |  | HTML iconHTML  

    The 21st IEEE International Workshop on Machine Learning for Signal Processing will be held in Beijing, China, on September 18–21, 2011. The workshop series is the major annual technical event of the IEEE Signal Processing Society's Technical Committee on Machine Learning for Signal Processing. This year the workshop is held in the National Laboratory of Pattern Recognition (NLPR), Institute of Automation, Chinese Academy of Sciences. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Organizing Committee

    Page(s): 1
    Save to Project icon | Request Permissions | PDF file iconPDF (261 KB)  
    Freely Available from IEEE
  • Program committee

    Page(s): 1 - 3
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (245 KB)  

    Provides a listing of current committee members. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Paper index

    Page(s): 1 - 7
    Save to Project icon | Request Permissions | PDF file iconPDF (416 KB)  
    Freely Available from IEEE
  • Protein subcellular localization prediction based on profile alignment and Gene Ontology

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (271 KB) |  | HTML iconHTML  

    The functions of proteins are closely related to their subcellular locations. Computational methods are required to replace the laborious and time-consuming experimental processes for proteomics research. This paper proposes combining homology-based profile alignment methods and functional-domain based Gene Ontology (GO) methods to predict the subcellular locations of proteins. The feature vectors constructed by these two methods are recognized by support vector machine (SVM) classifiers, and their scores are fused to enhance classification performance. The paper also investigates different approaches to constructing the GO vectors based on the GO terms returned from InterProScan. The results demonstrate that the GO methods are comparable to profile-alignment methods and overshadow those based on amino-acid compositions. Also, the fusion of these two methods can outperform the individual methods. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A sinusoidal audio and speech analysis/synthesis model based on improved EMD by adding pure tone

    Page(s): 1 - 5
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (331 KB) |  | HTML iconHTML  

    A multi-resolution speech and audio sinusoidal analysis/synthesis model based on an improved Empirical Mode Decomposition (EMD) is proposed in this paper. Because of the special filtering characteristic and superiority in dealing with non-stationary signal of EMD, a preprocessing module is adopted to classify the original signal by using the energy ratio and spectrum center of each Intrinsic Mode Function (IMF). A pure tone is added into original signal to extract the noise-like high frequency components without destroying the harmonics of signal. Then a multi-resolution Perceptual Weighted Matching Pursuit (PWMP) and frequency fine search method are adopted to estimate the sinusoidal parameters. Finally, objective measurements of perceived audio quality (PEAQ) show that this model can be effective for the audio synthesis. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Data representation and feature selection for colorimetric sensor arrays used as explosives detectors

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (142 KB) |  | HTML iconHTML  

    Within the framework of the strategic research project Xsense at the Technical University of Denmark, we are developing a colorimetric sensor array which can be useful for detection of explosives like DNT, TNT, HMX, RDX and TATP and identification of volatile organic compounds in the presence of water vapor in air. In order to analyze colorimetric sensors with statistical methods, the sensory output must be put into numerical form suitable for analysis. We present new ways of extracting features from a colorimetric sensor and determine the quality and robustness of these features using machine learning classifiers. Sensors, and in particular explosive sensors, must not only be able to classify explosives, they must also be able to measure the certainty of the classifier regarding the decision it has made. This means there is a need for classifiers that not only give a decision, but also give a posterior probability about the decision. We will compare K-nearest neighbor, artificial neural networks and sparse logistic regression for colorimetric sensor data analysis. Using the sparse solutions we perform feature selection and feature ranking and compare to Gram-Schmidt orthogonalization. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Efficient preference learning with pairwise continuous observations and Gaussian Processes

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (658 KB) |  | HTML iconHTML  

    Human preferences can effectively be elicited using pairwise comparisons and in this paper current state-of-the-art based on binary decisions is extended by a new paradigm which allows subjects to convey their degree of preference as a continuous but bounded response. For this purpose, a novel Beta-type likelihood is proposed and applied in a Bayesian regression framework using Gaussian Process priors. Posterior estimation and inference is performed using a Laplace approximation. The potential of the paradigm is demonstrated and discussed in terms of learning rates and robustness by evaluating the predictive performance under various noise conditions on a synthetic dataset. It is demonstrated that the learning rate of the novel paradigm is not only faster under ideal conditions, where continuous responses are naturally more informative than binary decisions, but also under adverse conditions where it seemingly preserves the robustness of the binary paradigm, suggesting that the new paradigm is robust to human inconsistency. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A novel method of diagnosing coronary heart disease by analysing ECG signals combined with motion activity

    Page(s): 1 - 5
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (317 KB) |  | HTML iconHTML  

    In this paper, we propose an effective method to automatically diagnose coronary heart disease by detecting ST segment episodes of ECG signals. To improve the diagnostic accuracy, we consider the motion activity of individual while monitoring ECG signals and we detect the motion activity of people through heart rate. Our method is based on clinical principle that ST segment depression is greater relative to heart rate (HR) in the recovery period compared with the exercise phase, which is stated in reference. Finally, the method is simulated by The Long-Term ST Database which has reference annotations about whether the person had coronary heart disease or not, with a diagnostic accuracy 80%. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • IVA for multi-subject FMRI analysis: A comparative study using a new simulation toolbox

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (420 KB) |  | HTML iconHTML  

    Joint blind source separation (JBSS) techniques have proven to be a natural solution for achieving source separation of multiple data sets. JBSS algorithms, such as independent vector analysis (IVA), are a promising alternative to independent component analysis (ICA) based approaches for the analysis of multi-subject functional magnetic resonance imaging (fMRI) data. Unlike ICA, little is known about the effectiveness of JBSS methods for fMRI analysis. In this paper, a new fMRI simulation toolbox (SimTB) is used to simulate multi-subject realistic fMRI datasets that include inter-subject variability. We study the performance of two JBSS algorithms representing two different approaches to the problem: (1) a recently proposed IVA algorithm combining second-order and higher-order statistics denoted by IVA-GL; and (2) a JBSS solution found by jointly diagonalizing cross-cumulant matrices denoted IVA-GJD. We compare these two JBSS algorithms with similar ICA algorithms implemented in the widely used group ICA for fMRI toolbox (GIFT). The results show that in addition to offering an effective solution for making group inferences, IVA algorithms provide superior performance in terms of capturing spatial inter-subject variability. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Deflation technique for neural spike sorting in multi-channel recordings

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (479 KB) |  | HTML iconHTML  

    We propose an ICA based algorithm for spike sorting in multi-channel neural recordings. In such context, the performance of ICA is known to be limited since the number of recording sites is much lower than the number of the neurons around. The algorithm uses an iterative application of ICA and a deflation technique in two nested loops. In each iteration of the external loop, the spiking activity of one neuron is singled out and then deflated from the recordings. The internal loop implements a sequence of ICA and spike detection for removing the noise and all the spikes that are not coming from the targeted neuron. We validate the performance of the algorithm on simulated data, but also on real simultaneous extracellular-intracellular recordings. The results show that the proposed algorithm performs significantly better than when only ICA is applied. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Heterogeneous mixture models using sparse representation features for applause and laugh detection

    Page(s): 1 - 5
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (304 KB) |  | HTML iconHTML  

    A novel and robust approach for applause and laugh detection is proposed based on sparse representation features and heterogeneous mixture models (hetMM). The projections of the noise robust sparse representations for audio signals computed by L1 - minimization are used as feature. We consider the classifiers based on heterogeneous mixture models (hetMM) which combine multiple different kinds of distributions, since in practice the data may come from multiple sources and it is often unclear what the most suitable distribution is. Experimental results show that method with hetMM has better results than using a single distribution type and gives comparable performances with Support Vector Machines (SVMs). View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Detection of playfield with shadow and its application to player tracking

    Page(s): 1 - 5
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (564 KB) |  | HTML iconHTML  

    Playfield detection is a key technology for content analysis in sports video, on which many semantic clue mining methods rely. However, shadow produced by substantial illumination change causes the general detection method fail and degenerate the performance of the following processing based on it. This paper presents a method for detecting playfield, which can find shadow region under the guidance of the intrinsic image proposed by Finlayson. Firstly, this method automatically finds the dominant color; secondly, according to the dominant color and the intrinsic image, it determines the playfield colors. At last, we apply it to player tracking in soccer video. Experimental results show that the proposed method can handle the problem brought by shadow and improve the performance of the application based on it. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Metric measurement from street view sequences with simple operator assistance and phase correlation based frame selection

    Page(s): 1 - 5
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (885 KB) |  | HTML iconHTML  

    This paper presents a metric measurement approach from sequences of images captured from a moving spherical camera without the need of additional equipment, such as laser scanners or motion detection units. The user assists the algorithms with simple inputs to facilitate the measurement process. The operator initially selects a keyframe that contains the object of interest that is to be measured. Next, a suitable pair is selected for this keyframe, automatically, using a novel phase correlation based approach proposed in this paper. Then, correspondence matching between these two images is performed using scale-invariant feature transform (SIFT) and these features are refined using RANdom SAmple Consensus (RANSAC) and information obtained from the phase correlation stage. As a last step conversion form the image domain to the 3D domain is performed. The user selects two corresponding point pairs in both frames, corresponding to the edges of the distance that is to be measured, and the metric distance between these two points is obtained. During this process, the height information of the camera with respect to the ground is used as basic reference to obtain metric results. Experimental results show that the proposed methods can provide metric measurements with up to 10% error. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • An asymptotic analysis of Bayesian state estimation in hidden Markov models

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (108 KB) |  | HTML iconHTML  

    Hidden Markov models are widely used for modeling underlying dynamics of sequence data. The accurate hidden state estimation is one of the central issues on practical application since the dynamics is described as a sequence of hidden states. However, while there are many studies on parameter estimation, mathematical properties of the hidden state estimation have not been clarified yet. The present paper analyzes the accuracy of a Bayesian hidden state estimation and shows that the dominant order of an error function depends on redundancy of states. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Compact and robust fisher descriptors for large-scale image retrieval

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (415 KB) |  | HTML iconHTML  

    Vector of locally aggregated descriptors (VLAD) has overcome the lossy quantization of bag-of-words model (BOW), but its dimensionality is high for direct use. We reduce the dimensionality of VLAD by a special coding scheme. First descriptors are clustered, and then linear discriminant analysis (LDA) is performed separately within each cluster. For different cluster, we allow different dimensionality but retain the same discriminant power, aiming at optimization of total dimensionality. Furthermore, we use each feature's nearest set of cluster centers as its expression bases, which is chosen using nearest neighbor distance ratio, so that the correspondence between a feature and its nearest set is more stable. The goal of the above scheme is to adapt the feature representation to distribution of feature classes in each cluster and distribution of cluster centers in feature space. Experiments demonstrate that our approach outperforms the state-of-the-art in computational complexity, accuracy, and robustness. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A new scatter-based multi-class support vector machine

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (313 KB) |  | HTML iconHTML  

    We provide a novel interpretation of the dual of support vector machines (SVMs) in terms of scatter with respect to class prototypes and their mean. As a key contribution, we extend this framework to multiple classes, providing a new joint Scatter SVM algorithm, at the level of its binary counterpart in the number of optimization variables. We identify the associated primal problem and develop a fast chunking-based optimizer. Promising results are reported, also compared to the state-of-the-art, at lower computational complexity. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Kernel entropy component analysis: New theory and semi-supervised learning

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (335 KB) |  | HTML iconHTML  

    A new theory for kernel entropy component analysis (kernel ECA) is developed, based on distribution dependent convolution operators, ensuring the validity of the method for any positive semi-definite kernel. Furthermore, a new semi-supervised kernel ECA classification method is derived with positive results compared to the state-of-the-art. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Active one-class learning by kernel density estimation

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (347 KB) |  | HTML iconHTML  

    Active learning has been a popular area of research in recent years. It can be used to improve the performance of learning tasks by asking the labels of unlabeled data from the user. In these methods, the goal is to achieve the highest possible accuracy gain while posing minimum queries to the user. The existing approaches for active learning have been mostly applicable to the traditional binary or multi-class classification problems. However, in many real-world situations, we encounter problems in which we have access only to samples of one class. These problems are known as one-class learning or outlier detection problems and the User relevance feedback in image retrieval systems is an example of such problems. In this paper, we propose an active learning method which uses only samples of one class. We use kernel density estimation as the baseline of one-class learning algorithm and then introduce some confidence criteria to select the best sample to be labeled by the user. The experimental results on real world and artificial datasets show that in the proposed method, the average gain in accuracy is increased significantly, compared to the popular random unlabeled sample selection strategy. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Large scale topic modeling made practical

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (465 KB) |  | HTML iconHTML  

    Topic models are of broad interest. They can be used for query expansion and result structuring in information retrieval and as an important component in services such as recommender systems and user adaptive advertising. In large scale applications both the size of the database (number of documents) and the size of the vocabulary can be significant challenges. Here we discuss two mechanisms that can make scalable solutions possible in the face of large document databases and large vocabularies. The first issue is addressed by a parallel distributed implementation, while the vocabulary problem is reduced by use of large and carefully curated term set. We demonstrate the performance of the proposed system and in the process break a previously claimed `world record' announced April 2010 both by speed and size of problem. We show that the use of a WordNet derived vocabulary can identify topics at par with a much larger case specific vocabulary. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Underdetermined convolutive blind source separation using a novel mixing matrix estimation and MMSE-based source estimation

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (520 KB) |  | HTML iconHTML  

    This paper considers underdetermined blind source separation of super-Gaussian signals that are convolutively mixed. The separation is performed in three stages. In the first stage, the mixing matrix in each frequency bin is estimated by the proposed single source detection and clustering (SSDC) algorithm. In the second stage, by assuming complex-valued super-Gaussian distribution, the sources are estimated by minimizing a mean-square-error (MSE) criterion. Special consideration is given to reduce computational load without compromising accuracy. In the last stage, the estimated sources in each frequency bin are aligned for recovery. In our simulations, the proposed algorithm outperformed conventional algorithm in terms of the mixing-error-ratio and the signal-to-distortion ratio. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Robust online estimation of the vanishing point for vehicle mounted cameras

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (1217 KB) |  | HTML iconHTML  

    For cameras mounted on a vehicle, the estimation of the vanishing point corresponding to the observed field of view is an important machine vision task necessary for a lot of applications, such as camera calibration and autonomous vehicle navigation. In this paper, a novel method for the estimation of the vanishing point corresponding to a particular camera orientation with respect to the vehicle is proposed. Robust features are first extracted and the motion of the vehicle is then used to estimate parallel trajectories by tracking these features. Thus, the proposed scheme does not rely on any man-made structures or pre-assumed gradients. The estimated trajectories are then processed to robustly estimate the vanishing point for the mounted camera for any given driving direction. Experimental results show that the proposed technique is able to robustly and accurately estimate the vanishing point for a variety of orientations of the camera mounting. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Gaussian process for human motion modeling: A comparative study

    Page(s): 1 - 6
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (335 KB) |  | HTML iconHTML  

    We evaluate recent Gaussian process (GP)-based manifold learning methods for human motion modeling, including our recently proposed joint gait and pose manifolds (JGPMs). Unlike most GP algorithms that involve either one latent variable or multiple independent variables in separate latent spaces, JGPMs define two variables jointly and explicitly in one latent space to represent a collection of gait data from different individuals. We develop a model validation technique to examine these GP-based algorithms in terms of their capability of motion interpolation, extrapolation, filtering, and recognition. Experimental results on both CMU Mocap and Brown HumanEva datasets show the superiority of JGPMs over existing GP algorithms for human motion modeling. View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.