Search

The common age-dependent West syndrome can be diagnosed accurately by electroencephalogram (EEG), but its pathogenesis and evolution remain unclear. Existing research mainly aims at the study of West seizure markers in time/frequency domain, while less literature uses a graph-theoretic approach to analyze changes among different brain regions. In this paper, the scalp EEG based functional connectivity (including Correlation, Coherence, Time Frequency Cross Mutual Information, Phase-Locking Value, Phase Lag Index, Weighted Phase Lag Index) and network topology parameters (including Clustering coefficient, Feature path length, Global efficiency, and Local efficiency) are comprehensively studied for the prognostic analysis of the West episode cycle. The scalp EEGs of 15 children with clinically diagnosed string spasticity seizures are used for prospective study, where the signal is divided into pre-seizure, seizure, and post-seizure states in 5 typical brain wave rhythm frequency bands (δ (1-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (30-80 Hz)) for functional connectivity analysis. The study shows that recurrent West seizures weaken connections between brain regions responsible for cognition and intelligence, while brain regions responsible for information synergy and visual reception have greater variability in connectivity during seizures. It is observed that the changes inβandγfrequency bands of the multiband brain network connectivity patterns calculated by Corr and WPLI can be preliminarily used as judgment of seizure cycle changes in West syndrome.

One-class classification achieves anomaly/outlier detection by exploiting the characteristics of target data. As a local similarity measure defined in kernel space, correntropy is generally more robust than the mean square error (MSE) based criterion in dealing with large outliers. In this paper, the maximum mixture correntropy criterion (MMCC) with multiple kernels are applied to the shallow and hierarchical one-class extreme learning machine (OC-ELM) to enhance the model robustness and learning speed. Experiments on benchmark UCI classification datasets, urban acoustic classification (UAC) dataset and 4 synthetic datasets are carried out to show the effectiveness and comparisons with several state-of-the-art methods are provided to demonstrate the superiority of the proposed algorithms.

Most multilayer Moore-Penrose inverse (MPI)-based neural networks, such as deep random vector functional link (RVFL), are structured with two separate stages: unsupervised feature encoding and supervised pattern classification. Once the unsupervised learning is finished, the latent encoding is fixed without supervised fine-tuning. However, in complex tasks such as handling the ImageNet dataset, there are often many more clues that can be directly encoded, while unsupervised learning, by definition, cannot know exactly what is useful for a certain task. There is a need to retrain the latent space representations in the supervised pattern classification stage to learn some clues that unsupervised learning has not yet been learned. In particular, the residual error in the output layer is pulled back to each hidden layer, and the parameters of the hidden layers are recalculated with MPI for more robust representations. In this article, a recomputation-based multilayer network using Moore-Penrose inverse (RML-MP) is developed. A sparse RML-MP (SRML-MP) model to boost the performance of RML-MP is then proposed. The experimental results with varying training samples (from 3k to 1.8 million) show that the proposed models provide higher Top-1 testing accuracy than most representation learning algorithms. For reproducibility, the source codes are available at https://github.com/W1AE/Retraining.

Soft robotic hands are inherently safer and more compliant in robot-environment interaction than rigid manipulators, but their flexibility and versatility still need improving. In this article, a gesture adaptive soft-rigid robotic hand is proposed. The robotic hand has three pneumatic two-segment fingers. Each finger segment is driven independently for flexible gesture adjustment to match up with different object shapes. The palm is constructed by a rigid skeleton driven by a soft pneumatic spring. It provides a firm support, large workspace, and independent force control for the fingers. Geometry model of the robotic hand is established, based on which a grasping gesture optimization algorithm is adopted. The fingers achieve optimal contact with objects by performing maximal curving similarity with the object outlines. Experiment shows that the soft-rigid robotic hand provides adaptive and reliable grasping for objects of different sizes, shapes, and materials with optimized gestures.

The extreme learning machine‐based autoencoder (ELM‐AE) has attracted a lot of attention due to its fast learning speed and promising representation capability. However, the existing ELM‐AE algorithms only reconstruct the original input and generally ignore the probability distribution of the data. The minimum error entropy (MEE), as an optimal criterion considering the distribution statistics of the data, is robust in handling non‐linear systems and non‐Gaussian noises. The MEE is equivalent to the minimisation of the Kullback–Leibaler divergence. Inspired by these advantages, a novel randomised AE is proposed by adopting the MEE criterion as the loss function in the ELM‐AE (in short, the MEE‐RAE) in this study. Instead of solving the output weight by the Moore–Penrose generalised inverse, the optimal output weight is obtained by the fixed‐point iteration method. Further, a quantised MEE (QMEE) is applied to reduce the computational complexity of. Simulations have shown that the QMEE‐RAE not only achieves superior generalisation performance but is also more robust to non‐Gaussian noises than the ELM‐AE.

Quaternion random neural network trained by extreme learning machine (Q-ELM) becomes attractive for its good learning capability and generalization performance in 3 or 4-dimensional (3/4-D) hypercomplex data learning. But how to determine the optimal network architecture is always challenging in Q-ELM. To this end, a novel error-minimization-based Q-ELM (QEM-ELM) that only needs to optimize the output weights of the newly added neuron is developed in this brief. On this basis, a dynamic network construction scheme is further extended on Q-ELM, leading to a novel DQ-ELM, where the hidden nodes can be dynamically recruited or deleted according to the significance to network performance. The network parameters can be optimized and the architecture can be self-adapted simultaneously. Simulation results on many benchmark datasets demonstrate that the proposed QEM-ELM and DQ-ELM achieve good generalization performance by preserving a compact network size.

Autoencoding is a vital branch of representation learning in deep neural networks (DNNs). The extreme learning machine-based autoencoder (ELM-AE) has been recently developed and has gained popularity for its fast learning speed and ease of implementation. However, the ELM-AE uses random hidden node parameters without tuning, which may generate meaningless encoded features. In this brief, we first propose a within-class scatter information constraint-based AE (WSI-AE) that minimizes both the reconstruction error and the within-class scatter of the encoded features. We then build stacked WSI-AEs into a one-class classification (OCC) algorithm based on the hierarchical regularized least-squared method. The effectiveness of our approach was experimentally demonstrated in comparisons with several state-of-the-art AEs and OCC algorithms. The evaluations were performed on several benchmark data sets.

Hierarchical neural network based one-class anomaly detection algorithms generally rely on stacked autoencoders (AEs) for feature learning. But existing AEs are not specifically designed to exploit the discriminative characteristics of target data in one-class classification (OCC), and thus may lead to poor generalization performance. In this brief, a novel randomized AE that imposes the constraint of the within-class scatter information is developed in feature learning. The correntropy criterion is applied to replace the mean square error criterion (MSE) to enhance the algorithm performance in outlier and noise rejection. The algorithm is further extended to kernel learning to improve its generalization capability. Experiments on benchmark datasets are carried out to show the effectiveness of the proposed algorithm.

Underground pipeline network suffers severe damages caused by construction machines used in the rapid urbanization development and construction, leading to profound impact to people’s life. An intelligent construction machinery classification (CMC) system thus becomes important to urban security and smart city engineering. Conventional methods are always problematic in one way or another. For instance, the one-dimensional speech features based recognition algorithms are usually less discriminative in the complex urban acoustic environment, and the two-dimensional spectrogram based convolution neural network (CNN) suffers a long model learning time and is less robust to the background noises. To address these deficiencies, a novel CMC system based on new hybrid acoustic features is presented. Two novel acoustic feature extraction methods are developed, where the first explores the concentrations of the Mel-frequency spectrogram (MFS) using the statistics of the binary MFS (SBMS) and the second characterizes the information entropy sequence feature of the binary MFS (IESF) based on the pulse-coupled neural network (PCNN) model. Then, the hybrid features of SBMS/IESF and the Linear Prediction Cepstral Coefficients (LPCC)/Mel-Frequency Cepstral Coefficients (MFCC) are further studied. A database consisting of 84,900 acoustic samples from 6 typical construction machines, horns of vehicles and other urban noises is recorded for experiment and performance validation. The comparisons with many state-of-the-art algorithms demonstrate the superiority of the proposed method.

Multi-source domain adaptation (MDA) aims to leverage knowledge from multiple source domains to improve the classification performance on target domains. Different degrees of distribution discrepancies between every two domains pose a huge challenge to MDA tasks. Most works focus on extracting features shared by all domains, which is critical but not enough to reduce distribution discrepancies. In this paper, we propose a method named as common-specific feature learning (CSFL). Constituting a framework of feature learning, CSFL explores a subspace where the combination of common and specific features makes learned representations comprehensive. Based on this framework, we conduct a metric learning method for learning a discriminative feature representation. Considering redundant information caused by source domains is likely to hurt the performance, we impose an effective low-rank constraint to remove the redundant information. Further, we adopt structure consistent constraint to preserve the local structure in each domain. CSFL has obtained about 1–5% improvement of mean accuracy, compared to the state-of-the-art shallow methods. Further, compared with 90.2% and 89.4% of the best baseline deep method, CSFL achieves mean accuracy of 90.8% and 89.7% on the Office-31 and ImageCLEF-DA datasets respectively. The encouraging results validate the effectiveness of our method.

Image segmentation is an important analysis tool in the field of computer vision. In this paper, on the basis of the traditional level set method, a novel segmentation model using generalized divergences is proposed. The main advantage of generalized divergences is their smooth connection performance among various kinds of well-known and frequently used fundamental divergences with one formula. Therefore, the discrepancy between two probability distributions of segmented image parts can be measured by generalized divergences. We also found a solution to determine the optimal divergence automatically for different images. Experimental results on a variety of synthetic and natural images are presented, which demonstrate the potential of the proposed method. Compared with the previous active contour models formulated to solve the same nonparametric statistical segmentation problem, our method performs better both qualitatively and quantitatively.

This paper presents WALLACE, a new framework of deep convolutional neural networks, which perform ConvNet's pyramidal feature hierarchy for weakly supervised learning. Most prior works rely on the image pyramid or network ensemble, which is both complicated and usually expensive. Instead, WALLACE is a more simple single-stage network that can predict objects present and location in an image without multiple rescale. Our model is trained efficiently using only global image-level labels, and it could generate meaningful multi-scale semantic feature maps by only one evaluation. Furthermore, a novel constrain-to-highlight loss is proposed to balances region selection among hierarchical feature maps, which additional improve model performance. Extensive experiments on object classification and weakly supervised pointwise object localization show that WALLACE achieves state-of-the-art results on the VOC 2007 and VOC 2012 benchmark without bells and whistles.

Employing prior information can greatly improve the segmentation result of many image segmentation problems. For example, a commonly used prior information is the shape of the object. In this paper, we introduce a different kind of prior information called the prior distribution. On the basis of non-parametric statistical active contour model, we add prior distribution energy to build a novel prior active contour model. During the convergence of contour curve, distribution difference between the inside and outside of the active contour is maximized while the distribution difference between the inside/outside of contour and the prior object/background is minimized. Furthermore, in order to improve the computation speed, a method to accelerate the computation speed is also proposed, which significantly relieves the burden of estimating probability density functions. As the experimental results suggest, satisfactory effects can be achieved in the segmentation of synthetic images and natural images via the our algorithm. Compared with the traditional non-parametric statistical active contour model without prior information, our method achieves a distinct improvement in both accuracy and computation efficiency.

One-class classification has been found attractive in many applications for its effectiveness in anomaly or outlier detection. Representative one-class classification algorithms include the one-class support vector machine (SVM), Naive Parzen density estimation, autoencoder (AE), etc. Recently, the one-class extreme learning machine (OC-ELM) has been developed for learning acceleration and performance enhancement. But existing one-class algorithms are generally less effective in complex and multi-class classifications. To alleviate the deficiency, a multilayer neural network based one-class classification with ELM (in short, as ML-OCELM) is developed in this paper. The stacked AEs are employed in ML-OCELM to exploit an effective feature representation for complex data. The effective kernel based learning framework is also investigated in the stacked AEs of ML-OCELM, leading to a multilayer kernel based OC-ELM (in short, as MK-OCELM). The MK-OCELM has advantages of less human-intervention parameters and good generalization performance. Experiments on 13 benchmark UCI classification datasets and a real application on urban acoustic classification (UAC) are carried out to show the superiority of the proposed ML-OCELM/MK-OCELM over the OC-ELM and several state-of-the-art algorithms.

El Niño and Southern Oscillation (ENSO) is closely related to a series of regional extreme climates, so robust long-term forecasting is of great significance for reducing economic losses caused by natural disasters. Here, we regard ENSO prediction as an unsupervised spatiotemporal prediction problem, and design a deep learning model called Dense Convolution-Long Short-Term Memory (DC-LSTM). For a more sufficient training model, we will also add historical simulation data to the training set. The experimental results show that DC-LSTM is more suitable for the prediction of a large region and a single factor. During the 1994–2010 verification period, the all-season correlation skill of the Nino3.4 index of the DC-LSTM is higher than that of the current dynamic model and regression neural network, and it can provide effective forecasts for lead times of up to 20 months. Therefore, DC-LSTM can be used as a powerful tool for predicting ENSO events.

Accurate classification of the children's epilepsy syndrome is vital to the diagnosis and treatment of epilepsy. But existing literature mainly focuses on seizure detection and few attention has been paid to the children's epilepsy syndrome classification. In this paper, we present a study on the classification of two most common epilepsy syndromes: the benign childhood epilepsy with centro-temporal spikes (BECT) and the infantile spasms (also known as the WEST syndrome), recorded from the Children's Hospital, Zhejiang University School of Medicine (CHZU). A novel feature fusion model based on the deep transfer learning and the conventional time-frequency representation of the scalp electroencephalogram (EEG) is developed for the epilepsy syndrome characterization. A fully connected network is constructed for the feature learning and syndrome classification. Experiments on the CHZU database show that the proposed algorithm can offer an average of 92.35% classification accuracy on the BECT and WEST syndromes and their corresponding normal cases.

The imbalanced data classification attracts increasing attention in the past years due to the continuous expansion of data available in many areas, such as biomedical engineering, surveillance, and computer vision. Learning from imbalanced data is challenging as most standard algorithms fail to properly represent the inherent complex characteristics of the data distribution. As an emerging technology, the extreme learning machine (ELM) and its variants, including the weighted ELM (WELM) and the boosting weighted ELM (BWELM), have been recently developed for the classification of imbalanced data. However, the WELM suffers the following deficiencies: (i) the sample weight matrix is manually chosen and fixed during the learning phase; (ii) the representation capability, namely the capability to extract features or useful information from the original data, is insufficiently explored due to its shallow structure. The BWELM employs the AdaBoost algorithm to optimize the sample weights. But the representation capability is still restricted by the shallow structure. To alleviate these deficiencies, we propose a novel deep weighted ELM (DWELM) algorithm for imbalanced data classification in this paper. An enhanced stacked multilayer deep representation network trained with the ELM (EH-DrELM) is first proposed to improve the representation capability, and a fast AdaBoost algorithm for imbalanced multiclass data (AdaBoost-ID) is developed to optimize the sample weights. Then, the novel DWELM for the imbalance learning is obtained by combining the above two algorithms. Experimental results on nine imbalanced binary-class datasets, nine imbalanced multiclass datasets, and five large benchmark datasets (three for multiclass and two for binary-class) show that the proposed DWELM achieves a better performance than the WELM and BWELM, as well as several state-of-the-art multilayer network-based learning algorithms.

Fear of heights (FoH) analysis and its association to physiological signals can better help understand people’s emotion and quantify human’s behavior, which have been found important in many applications, such as disease analysis, affective computing, etc. Existing studies are mainly on how to alleviate FoH while little literature on FoH analysis has been reported in the past. In this paper, we present the studies of correlation on FoH to multi-modal physiological signals in the virtual reality (VR). To stimulate the FoH to participants, 4 types of VR scenarios that consist of the virtual scene of the VR game “Richie’s plank experience” and the realistic stimulus of hitting by basketball are adopted in the experiment. The synchronized eye movement (EMO), pupil, and electrocardiogram (ECG) of 17 healthy subjects with an even mix of men and women are recorded for FoH analysis. The observations on the FoH analysis carried out in the paper are two-fold: (1) The recognition ability of multi-modal physiological signals for HoF has been evaluated using ReliefF feature selection algorithm. The results show that pupil feature based model generally achieves the best classification performance for FoH. Multiple features consisted of pupil diameter, power spectral density (PSD) of EMO, pupil, ECG, and mean value of EMO can effectively quantify the variability. (2) The FoH analysis model built on the correlation analysis algorithms combined with multi-modal feature-level fusion and strategy-level fusion methods can effectively overcome the drawbacks of conventional statistical models.

Underground pipelines suffered severe external breakage caused by excavation devices due to arbitral road excavation. Acoustic signal-based recognition has recently shown effectiveness in underground pipeline network surveillance. However, merely relying on recognition may lead to a high false alarm rate. The reason is that underground pipelines are generally paved along a fixed direction and excavations out of the region also trigger the surveillance system. To enhance the reliability of the surveillance system, the direction-of-arrival (DOA) estimation of target sources is combined into the recognition algorithm to reduce false detections in this paper. Two hybrid recognition algorithms are developed. The first one employs extreme learning machine (ELM) for acoustic recognition followed by a focusing matrix-based multiple signal classification algorithm (ELM-MUSIC) for DOA estimation. The second introduces a decision matrix (DM) to characterize the statistic distribution of results obtained by ELM-MUSIC. Real acoustic signals collected by a cross-layer sensor array are conducted for performance comparison. Four representative excavation devices working in a metro construction site are used to generate the signal. Multiple scenarios of the experiments are designed. Comparisons show that the proposed ELM-MUSIC and DM algorithms outperform the conventional focusing matrix based MUSIC (F-MUSIC). In addition, the improved DM method is capable of localizing multiple devices working in order. Two hybrid acoustic signal recognition and source direction estimation algorithms are developed for excavation device classification in this paper. The novel recognition combining DOA estimation scheme can work efficiently for underground pipeline network protection in the real-world complex environment.

One-class classification (OCC) has attracted a lot of attentions for its strong applicability of outlier/anomaly detection in real-world applications. Among them, the mixed model based on the autoencoder (AE) is the commonly used OCC architecture. The recent random neural network based extreme learning machine (ELM) AE (ELM-AE) has been extended to OCC. But no constraints have been applied to the encoded feature in ELM-AE and the random parameters used in ELM-AE may result in meaningless features. In this paper, we propose a novel within-class scatter constraint based randomized AE (WSC-RAE) by introducing a constraint on the distribution of the encoded feature to restrict the solution space and thus avoid the meaningless features. The proposed WSC-RAE is then embedded into a multilayer one-class ELM (ML-OCELM) framework for OCC. Experiments on 17 benchmark datasets are carried out to demonstrate the superiority of WSC-RAE.

One of the salient features of the extreme learning machine (ELM) is its fast learning speed. However, in a big data environment, the ELM still suffers from an overly heavy computational load due to the high dimensionality and the large amount of data. Using the alternating direction method of multipliers (ADMM), a convex model fitting problem can be split into a set of concurrently executable subproblems, each with just a subset of model coefficients. By maximally splitting across the coefficients and incorporating a novel relaxation technique, a maximally split and relaxed ADMM (MS-RADMM), along with a scalarwise implementation, is developed for the regularized ELM (RELM). The convergence conditions and the convergence rate of the MS-RADMM are established, which exhibits linear convergence with a smaller convergence ratio than the unrelaxed maximally split ADMM. The optimal parameter values of the MS-RADMM are obtained and a fast parameter selection scheme is provided. Experiments on ten benchmark classification data sets are conducted, the results of which demonstrate the fast convergence and parallelism of the MS-RADMM. Complexity comparisons with the matrix-inversion-based method in terms of the numbers of multiplication and addition operations, the computation time and the number of memory cells are provided for performance evaluation of the MS-RADMM.

Along with the explosive growing of data, semi-supervised learning attracts increasing attention in the past years due to its powerful capability in labeling unlabeled data and knowledge mining. As an emerging method, the semi-supervised extreme learning machine (SSELM), that builds on ELM, has been developed for data classification and shown superiorities in learning efficiency and accuracy. However, the optimization of SSELM as well as most of the other ELMs is generally based on the mean square error (MSE) criterion, which has been shown less effective in dealing with non-Gaussian noises. In this paper, a robust regularized correntropy criterion based SSELM (RC-SSELM) is developed. The optimization of the output weight matrix of RC-SSELM is derived by the fixed-point iteration based approach. A convergent analysis of the proposed RC-SSELM is presented based on the half-quadratic optimization technique. Experimental results on 4 synthetic datasets and 13 benchmark UCI datasets are provided to show the superiorities of the proposed RC-SSELM over SSELM and other state-of-the-art methods.

Research on the virtual machining environment is one of the key parts of virtual manufacturing technology. According to the features of virtual turning, this paper proposes a simplified Whitted lighting model based on the analysis of Phong and other local illumination models. Moreover, to obtain natural lighting effects, a local ray tracing algorithm is given to calculate the light intensity of every position during the course of simulation. This method cannot simulate the machining environment accurately, but also can reduce the calculation time. Further, to obtain a more realistic machining scene, the paper deduces the transfer matrix from the chip coordinate to the feeding coordinate so that the chip model can be integrated into the whole machining scene. Finally, an example of a virtual cutting scene is shown to demonstrate the effects of the global illumination model and chip simulation. The data needed for the construction of the scene and for chip simulation is obtained from a C616 lathe. This study plays an important role in richening virtual manufacturing theory and promoting the development of advanced manufacturing technology.

The driving fatigue related human vocal organ changes can influence the strange atrractor's trajectories in phase space under a speech nonlinear dynamic model and correspondingly affect the chaos and fractal features of speech signal. Therefore, this paper proposes a method for detecting driver fatigue using nonlinear speech processing techniques combined with fuzzy support vector machine (SVM). First, we reconstructed the speech signal in phase space to build a speech nonlinear dynamic model, extracted and analyzed the changes of several nonlinear dynamic features such as the largest Lyapunov exponent, fractal dimension and approximate entropy when the degree of human fatigue changes. Then during the designing process of the multi-feature fatigue classifier, fuzzy clustering concept was introduced to the traditional SVM classifier, and the fuzzy membership degrees of the training speech samples are computed by a novel proposed membership function to overcome the sensitivity to noise and outliers. After training this fuzzy SVM, a multi-feature fusion classifier was established for fatigue recognition of drivers' speech samples. The experimental results confirm the feasibility and the effectiveness of this proposed method, and it seems to be promising in the application of driver fatigue detection.