Qing Wu

A robust generalization of isomap for new data

Most of existing nonlinear dimensionality reduction algorithms, such as isomap, LEE, Laplacian Eigenmaps, SPE and so on, do not provide a simple generalization to discover the low-dimensional embedding for new data points. In this paper, we present a robust extension for isomap to efficiently map new samples into the low-dimensional space. This generalization permits one to apply a trained model to new data points without having to recompute eigenvectors and can effectively treat data with noise. Two methods are used to estimate the geodesic distances between new data points and training points. Experimental results demonstrate that the proposed algorithm is effective.

Classify the number of EEG current sources using support vector machines

The classifier based on support vector machines (SVMs) has had successful applications in many fields for its simple structure and excellent learning performance. In this paper we apply such classifiers to the EEG (electroencephalogram) data and use them to determine the number of EEG current sources according to the scalp potentials. Experimental results indicate that SVM classifiers are an effective and promising approach for this task.

EEG source localization of ERP based on multidimensional support vector regression approach

A new integrated multi-method system is presented to estimate the location and moment of equivalent current dipole sources of event-related potentials (ERP). In order to handle the large-scale high dimension problems efficiently and quickly, the ISOMAP algorithm was used to find the low dimensional manifolds from recorded EEG. Then, based on reduced dimension data, multidimensional support vector regression (MSVR) with similar iterative re-weight least square (IRWLS) was used to discover the relationship between the observation potentials on the scalp and the internal sources within the brain. In our experiments, the two current dipole sources with four-shell concentric sphere model were reconstructed. Our experiments demonstrate that MSVR based on the support vector machine can obtain more robust estimations for EEG source localization problem.

A fast genetic algorithm for solving the maximum clique problem

Aiming at the defects of Genetic Algorithm (GA) for solving the Maximum Clique Problem (MCP) in more complicated, long-running and poor generality, a fast genetic algorithm (FGA) is proposed in this paper. A new chromosome repair method on the degree, elitist selection based on random repairing, uniform crossover and inversion mutation are adopted in the new algorithm. These components can speed up the search and effectively prevent the algorithm from trapping into the local optimum. The algorithm was tested on DIMACS benchmark graphs. Experimental results show that FGA has better performance and high generality.

The Method of Multidimensional Support Vector Regression for Moving Dipole Localization of Face Expression

Brain signal source localization is a process of inverse calculation from electroencephalogram (EEG) signal. A new method of Multidimensional Support Vector Regression (MSVR) with similar iterative re-weight least square (IRWLS) is firstly used in source localization of face expression. In order to discover the relationship between sensor information and internal source in the brain, the moving dipole with four-shell concentric sphere model was reconstructed. Its location parameters and components were fitted in a series of time points. EEG signals of face expression were adopted in our experiments. Satisfactory results demonstrate that MSVR based on the support vector machine can obtain more robust estimations for EEG inverse problem.

An Incremental Algorithm Based on K Nearest Neighbor Projection for Nonlinear Dimensionality Reduction

Recently, there are several algorithms to perform dimensionality reduction on low-dimensional nonlinear manifolds embedded in a high-dimensional space, such as ISOMAP, LLE, Laplacian eigenmaps, SPE and so on. Most of these techniques work in batch mode. In this paper, we present an incremental nonlinear dimensionality reduction algorithm based on the k nearest neighbor projection. The method can effectively map new data into the low-dimensional space by building a locally linear transformation model between the original space and the embedded space. Moreover, the algorithm can treat data set with noise. Experiments show that the algorithm proposed is effective and robust

Application of wavelet network combined with nonlinear dimensionality reduction on the neural dipole localization

A wavelet network (WN) method is presented in this paper, which can be used to estimate the location and moment of an equivalent current dipole source using reduced-dimension data from the original measurement electroencephalography (EEG). In order to handle the large-scale high dimension problems efficiently and provide a real-time EEG dipole source localizer, the ISOMAP algorithm is firstly used to find the low dimensional manifolds from high dimensional EEG signal. Then, a WN is employed to discover the relationship between the observation potentials on the scalp and the internal sources within the brain. In our simulation experiments, satisfactory results are obtained.

A multi-class classifying algorithm based on nonlinear dimensionality reduction and support vector machines

Many problems in pattern classifications involve some form of dimensionality reduction. ISOMAP is a representative nonlinear dimensionality reduction algorithm, which can discover low dimensional manifolds from high dimensional data. To speed ISOMAP and decrease the dependency to the neighborhood size, we propose an improved algorithm. It can automatically select a proper neighborhood size and an appropriate landmark set according to a stress function. A multi-class classifier with high efficiency is obtained through combining the improved ISOMAP with SVM. Experiments show that the classifier presented is effective in fingerprint classifications.

Support vector machine method using in EEG signals study of epileptic spike

Support vector machine (SVM) is a new method of Machine Learning. SVR algorithms are normally only used for single-output systems now. Several SVR models were evaluated to identify one appropriate for multi-input multi-output systems, which require a much more complex control system. Based on good understanding of the SVM theory and algorithm, our studies discussed the multi-dimensional support vector regression (MSVR) and improved its algorithms. Electroencephalogram (EEG) source localization is well known as an import inverse problem of electrophysiology. In order to improve the accuracy of inverse calculation from EEG signal, MSVR is first applied in inverse problems, it has the advantages of simpler operation, faster convergence and better effect compared with single output SVR.

Research on the Neural Dipole Localization Using a Method Combining SVM with Nonlinear Dimensionality Reduction

Electroencephalogram (EEG) source localization is well known as an import inverse problem of electrophysiology. In order to improve the accuracy of inverse calculation from EEG signal, a new method combining multidimensional SVR with nonlinear dimensionality reduction was proposed. In our study, the ISOMAP algorithm was firstly used to find the low dimensional manifolds from high dimensional EEG signal. Then, a new method of Multidimensional Support Vector Regression (MSVR) with similar iterative re-weight least square (IRWLS) was applied to discover the parameters of EEG signals. In our experiments, EEG signals of epileptic spike were adopted as the objects. The satisfactory results were obtained.