Extracting fine-grained features such as styles from unlabeled data is crucial for data analysis. Unsupervised methods such as variational autoencoders (VAEs) can extract styles that are usually mixed with other features. Conditional VAEs (CVAEs) can isolate styles using class labels; however, there are no established methods to extract only styles using unlabeled data. In this paper, we propose a CVAE-based method that extracts style features using only unlabeled data. The proposed model consists of a contrastive learning (CL) part that extracts style-independent features and a CVAE part that extracts style features. The CL model learns representations independent of data augmentation, which can be viewed as a perturbation in styles, in a self-supervised manner. Considering the style-independent features from the pretrained CL model as a condition, the CVAE learns to extract only styles. Additionally, we introduce a constraint based on mutual information between the CL and VAE features to prevent the CVAE from ignoring the condition. Experiments conducted using two simple datasets, MNIST and an original dataset based on Google Fonts, demonstrate that the proposed method can efficiently extract style features. Further experiments using real-world natural image datasets were also conducted to illustrate the method’s extendability.

In numerous applications, binary reactions or event counts are observed and stored within high-order tensors. Tensor decompositions (TDs) serve as a powerful tool to handle such high-dimensional and sparse data. However, many traditional TDs are explicitly or implicitly designed based on the Gaussian distribution, which is unsuitable for discrete data. Moreover, most TDs rely on predefined multi-linear structures, such as CP and Tucker formats. Therefore, they may not be effective enough to handle complex real-world datasets. To address these issues, we propose ENTED, an Efficient Nonparametric TEnsor Decomposition for binary and count tensors. Specifically, we first employ a nonparametric Gaussian process (GP) to replace traditional multi-linear structures. Next, we utilize the Pólya-Gamma augmentation which provides a unified framework to establish conjugate models for binary and count distributions. Finally, to address the computational issue of GPs, we enhance the model by incorporating sparse orthogonal variational inference of inducing points, which offers a more effective covariance approximation within GPs and stochastic natural gradient updates for nonparametric models. We evaluate our model on several real-world tensor completion tasks, considering binary and count datasets. The results manifest both better performance and computational advantages of the proposed model.

Manual seizure detection from long recordings of the electroencephalogram (EEG) is a tiring, tedious, and error-prone process. It also requires experienced practitioners to detect seizure events precisely. This paper has proposed a novel method to detect epileptic seizures from long-EEG recordings using some state-of-the-art anomaly detectors and artifact rejection techniques. This method has two stages: 1) pre-screened EEGs for potential seizure detection and 2) seizure detection by artifact removal. For Stage 1, we have experimented with six methods of well-known anomaly detection. In Stage 2, we have proposed some artifact removal techniques to separate artifacts from seizure events based on the features calculated from the statistical properties of EEG. The proposed method has been evaluated with a private EEG dataset from Juntendo University Medical School, Japan. The proposed approach performed quite well with sensitivity (86–89%), specificity (92–94%), precision (20–22%), and seizure event detection ratio (89–91%). The promising results of the method encourage further improvement in the near future.

The deep neural networks are known to be vulnerable to well-designed adversarial attacks. The most successful defense technique based on adversarial training (AT) can achieve optimal robustness against particular attacks but cannot generalize well to unseen attacks. Another effective defense technique based on adversarial purification (AP) can enhance generalization but cannot achieve optimal robustness. Meanwhile, both methods share one common limitation on the degraded standard accuracy. To mitigate these issues, we propose a novel pipeline to acquire the robust purifier model, named Adversarial Training on Purification (AToP), which comprises two components: perturbation destruction by random transforms (RT) and purifier model fine-tuned (FT) by adversarial loss. RT is essential to avoid overlearning to known attacks, resulting in the robustness generalization to unseen attacks, and FT is essential for the improvement of robustness. To evaluate our method in an efficient and scalable way, we conduct extensive experiments on CIFAR-10, CIFAR-100, and ImageNette to demonstrate that our method achieves optimal robustness and exhibits generalization ability against unseen attacks.

Epileptic seizure detection from long recordings of scalp electroencephalography (EEG) is a challenging task owing to their unpredictability in nature with the inclusion of noise, artifacts and subject dependency. We hypothesize that selection of training EEG data plays important role in the model performance. Thus, we introduced an active learning based training data selection and modification method with a Riemannian geometry, centroid alignment, tangent space mapping and a support vector machine classifier. The proposed method focuses on the strategy of a patient-specific training approach with a patient-independent evaluation. Both the spatial and temporal correlations between and within the EEG channels are taken into the consideration. The benchmark CHB-MIT dataset is used to evaluate our model. For our model, we reported an average sensitivity of 86.8%, specificity of 85.4%, and accuracy of 86.1% with a receiver operating characteristics-area under the curve (ROC-AUC) of 93.1%, which has encouraged us to further improve its performance.

This paper reports on a novel invasive brain–computer interface (BCI) paradigm that has successfully reconstructed spoken sentences from invasive electrocorticogram (ECoG) signals using deep-neural-network-based encoders and a pre-trained neural vocoder. We recorded ECoG signals while 13 participants were speaking short sentences. Our BCI could map the ECoG recording to the log-mel spectrograms of the spoken sentences using a bidirectional long short-term memory (BLSTM) or a Transformer. The estimated log-mel spectrograms were used in Parallel WaveGAN to synthesize speech waveforms. An evaluation of the model performance revealed that the Transformer model significantly outperformed (Wilcoxon signed-rank test, p < 0.001) the BLSTM in terms of mean square error loss and Pearson correlation.

Tensor decompositions (TDs) serve as a powerful tool for analyzing multiway data. Traditional TDs incorporate prior knowledge about the data into the model, such as a directed generative process from latent factors to observations. In practice, selecting proper structural or distributional assumptions beforehand is crucial for obtaining a promising TD representation. However, since such prior knowledge is typically unavailable in real-world applications, choosing an appropriate TD model can be challenging. This paper aims to address this issue by introducing a flexible TD framework that discards the structural and distributional assumptions, in order to learn as much information from the data. Specifically, we construct a TD model that captures the joint probability of the data and latent tensor factors through a deep energy-based model (EBM). Neural networks are then employed to parameterize the joint energy function of tensor factors and tensor entries. The flexibility of EBM and neural networks enables the learning of underlying structures and distributions. In addition, by designing the energy function, our model unifies the learning process of different types of tensors, such as static tensors and dynamic tensors with time stamps. The resulting model presents a doubly intractable nature due to the presence of latent tensor factors and the unnormalized probability function. To efficiently train the model, we derive a variational upper bound of the conditional noise-contrastive estimation objective that learns the unnormalized joint probability by distinguishing data from conditional noises. We show advantages of our model on both synthetic and several real-world datasets.

Automated identification of epileptiform discharges in electroencephalograms (EEG) for the diagnosis of epilepsy can mitigate the burden of manual searches. Recent effective methods based on machine learning–based classification have used detection of candidate waveforms with signal processing and pattern matching as preprocessing, and this method can determine the overall performance. This paper thus considers a scenario where candidates are not detected; that is, we propose a recurrent neural network (RNN)–based self-attention model that can be fitted from the EEG segments generated without spike candidates being detected. In comparison with the state-of-the-art machine learning models that can be applied to EEG classification (LightGBM and EEGNet), the proposed model achieved higher performance (average accuracy: 90.2%). This result strongly suggests that the selfattention mechanism is suitable to automated identification of the epileptiform discharge in the EEG.

Invasive brain–machine interfaces (BMIs) are a promising neurotechnological venture for achieving direct speech communication from a human brain, but it faces many challenges. In this paper, we measured the invasive electrocorticogram (ECoG) signals from seven participating epilepsy patients as they spoke a sentence consisting of multiple phrases. A Transformer encoder was incorporated into a "sequence-to-sequence" model to decode spoken sentences from the ECoG. The decoding test revealed that the use of the Transformer model achieved a minimum phrase error rate (PER) of 16.4%, and the median (±standard deviation) across seven participants was 31.3% (±10.0%). Moreover, the proposed model with the Transformer achieved significantly better decoding accuracy than a conventional long short-term memory model.

Deep learning techniques are able to extract the characteristics of temporal signals to study their patterns and diagnose diseases such as essential tremor. However, these techniques require a large amount of data to train the neural network and achieve good results, and the more data the network has, the more accurate the final model implemented will be. This work proposes the use of data augmentation techniques to improve the accuracy of a Long short-term memory system in the diagnosis of essential tremor. For this purpose, the Empirical Modal Decomposition method will be used to decompose the original temporal signals collected from control subjects and patients with essential tremor. The time series obtained from the decomposition, covering different frequency ranges, will be randomly shuffled and combined to generate new artificial samples for each group. Then, both the generated artificial samples and part of the real samples will be used to train the LSTM network, and the remaining original samples will be used to test the model. Experimental results demonstrate the capability of the proposed method, increasing the classifier accuracy from 83.2% to almost 93% when artificial samples are used.

To cope with the lack of highly skilled professionals, machine learning with proper signal processing is key for establishing automated diagnostic-aid technologies with which to conduct epileptic electroencephalogram (EEG) testing. In particular, frequency filtering with the appropriate passbands is essential for enhancing the biomarkers-such as epileptic spike waves-that are noted in the EEG. This paper introduces a novel class of neural networks (NNs) that have a bank of linear-phase finite impulse response filters at the first layer as a preprocessor that can behave as bandpass filters that extract biomarkers without destroying waveforms because of a linear-phase condition. Besides, the parameters of the filters are also data-driven. The proposed NNs were trained with a large amount of clinical EEG data, including 15 833 epileptic spike waveforms recorded from 50 patients, and their labels were annotated by specialists. In the experiments, we compared three scenarios for the first layer: no preprocessing, discrete wavelet transform, and the proposed data-driven filters. The experimental results show that the trained data-driven filter bank with supervised learning behaves like multiple bandpass filters. In particular, the trained filter passed a frequency band of approximately 10-30 Hz. Moreover, the proposed method detected epileptic spikes, with the area under the receiver operating characteristic curve of 0.967 in the mean of 50 intersubject validations.

Latent factor models are canonical tools to learn low-dimensional and linear embedding of original data. Traditional latent factor models are based on low-rank matrix factorization of covariance matrices. However, for higher-order data with multiple modes, i.e., tensors, this simple treatment fails to take into account the mode-specific relations. This ignorance leads to inefficiency in analysis of complex structures as well as poor data compression ability. In this paper, unlike covariance matrices, we investigate high-order covariance tensor directly by exploiting tensor ring (TR) format and propose the Bayesian TR latent factor model, which can represent complex multi-linear correlations and achieves efficient data compression. To overcome the difficulty of finding the optimal TR-ranks and simultaneously imposing sparsity on loading coefficients, a multiplicative Gamma process (MGP) prior is adopted to automatically infer the ranks and obtain sparsity. Then, we establish an efficient parameter-expanded EM algorithm to learn the maximum a posteriori (MAP) estimate of model parameters. Finally, we evaluate our model on covariance estimation, latent factor learning and image inpainting problems.

Epilepsy is a chronic disease caused by sudden abnormal discharges of neurons which lead to transient brain disfunction. For those patients who need surgery, the recognition of the epileptic foci is one of the most important steps. In this paper, a combination of EEG signals processed by Short-time Fourier Transform and processed by Continuous Wavelet Transform is introduced. A deep learning model with an accuracy of 91.3% trained by Convolution Neural Network has been obtained. The experiment uses a methodology that has not been widely tested so far, and it also demonstrated the practical values by its relatively small time consumption and computing power requirements meanwhile ensuring high accuracy. Furthermore, it proves the feasibility of improving accuracy by combining different feature extraction methods.

We propose ScalpNet: A deep neural network to detect spatiotemporal abnormal intervals from EEGs of epilepsy patients. Since the number of trained clinicians is very limited, it is very crucial to establish automatic detection of abnormal signals caused by epilepsy from EEGs. We build a convolutional neural network detecting spatiotemporal intervals that will be abnormal based on the fact that peaky EEG signals can be observed not only in the electrode close to the focal region but those in the surrounding regions. In the experiments with a real dataset, our proposed ScalpNet presents higher classification accuracy than existing machine learning methods, including a convolutional neural network performed by channel-by-channel.

Epileptic focus localization in patients with epileptic seizures is essential when surgery is needed. Recent studies show that this can be done automatically using machine learning approaches. However, well-designed feature extraction methods are often computationally demanding, requiring a large amount of data labeled by physicians, which is time consuming and impractical. In this paper, we firstly introduce a one-dimensional convolutional neural network (1D-CNN) model for epileptic seizure focus detection which avoids the manual, time-consuming feature extraction Moreover, to reduce the necessary number of training samples, we introduce an approach for data augmentation. The experimental results demonstrate the efficiency of the proposed method, with a nearly 3% improvement in performance using the data enhancement method compared to the best result obtained using the traditional feature extraction method.

Electroencephalography (EEG) is considered as a potential tool for diagnosis of epilepsy in clinical applications. Epileptic seizures occur irregularly and unpredictably. Its automatic detection in EEG recordings is highly demanding. In this work, multiband features are used to detect seizure with feedforward neural network (FfNN). The EEG signal is segmented into epochs of short duration and each epoch is decomposed into a number of subbands using discrete wavelet transform (DWT). Three features namely ellipse area of second-order difference plot, coefficient of variation and fluctuation index are computed from each subband signal. The features obtained from all subbands are combined to construct the feature vector. The FfNN is trained using the derived feature vector and seizure detection is performed with test data. The experiment is performed with publicly available dataset to evaluate the performance of the proposed method. The experimental results show the superiority of this method compared to the recently developed algorithms.

The major challenge in Brain Computer Interface (BCI) is to obtain reliable classification accuracy of motor imagery (MI) task. This paper mainly focuses on unsupervised feature selection for electroencephalography (EEG) classification leading to BCI implementation. The multichannel EEG signal is decomposed into a number of subband signals. The features are extracted from each subband by applying spatial filtering technique. The features are combined into a common feature space to represent the effective event MI classification. It may inevitably include some irrelevant features yielding the increase of dimension and mislead the classification system. The unsupervised discriminative feature selection (UDFS) is employed here to select the subset of extracted features. It effectively selects the dominant features to improve classification accuracy of motor imagery task acquired by EEG signals. The classification of MI tasks is performed by support vector machine. The performance of the proposed method is evaluated using publicly available dataset obtained from BCI Competition III (IVA). The experimental results show that the performance of this method is better than that of the recently developed algorithms.

Epilepsy is a chronic disorder that causes unprovoked, recurrent-seizures. Characteristic spikes are often observed in the electroencephalogram (EEG) of epileptic patients in order to diagnose the disorder. Several methods have been investigated to automatically detect such spikes. The most common methods employ sub-band decomposition with discrete wavelet transform (DWT) or other filters to preprocess the EEG data before feeding it into a machine learning model. This paper introduces a fully data-driven method that automatically determines EEG frequency bands of interest. The raw signal is fed into a convolutional layer to detect suitable frequency bands, followed by a feedforward convolutional neural network (CNN) model or recurrent neural network (RNN) models for epileptic spike and non-spike classification. Fitting data of six patients, annotated by an epilepsy specialist, resulted in a convolutional layer with a frequency characteristic similar to bandpass filters. This result strongly justifies limiting the bandwidth of a signal, as done in previous studies. Moreover, results of the cross-subject validation indicate that a classical support vector machine with fixed preprocessing achieves comparable performance in the classification with fully data-driven models.

Multiple stimulus coding plays an important role in a steady-state visual evoked potential (SSVEP)-based brain-computer interface (BCI). In conventional SSVEP-based BCIs, multiple visual stimuli are modulated with different properties such as frequencies and/or phases. However, the number of properties that can be assigned to visual stimuli rendered on a computer monitor is always limited by its refresh rate, leading to a system with limited commands or functions. To alleviate this issue, this study proposes a novel waveform-based stimulus coding method that uses modulation waveforms to differentiate resulting SSVEPs. In this paper, the discriminability of 12-class SSVEPs modulated by three types of waveforms (i.e., rectangle, sinusoidal and triangle waveforms) and four frequencies (i.e., 12, 13, 14, and 15 Hz) was investigated by computing its classification accuracy. The results showed the SSVEPs modulated by different waveforms can be successfully distinguished when using the state-of-the-art canonical correlation analysis (CCA)-based method with an average accuracy of 92.31%. This result suggests that the proposed method has great potential to significantly increase the number of functions in an SSVEP-based BCI system.

This paper establishes an adaptive update method for the Gaussian kernel parameters in the application to the kernel adaptive filtering (KAF). In this method, the kernel parameters are all adaptive and data-driven, although they should be given or estimated by cross-validation. In terms of the Gaussian KAF, every input sample or signal has its own width and center, which are updated at each iteration based on the proposed least-square-type rules to minimize the estimation error. In particular, the proposed update rule keeps the width in the manifold of the positive numbers. Together with the l1- regularized least squares, the overall KAF algorithm can avoid the overfitting and the increase of dimensionality. Experimental results support the validity of the method.

This paper addresses the problem of finding optimal sensor placement, i.e., determining F sensor positions from N possible locations. We propose a sensor selection method based on the localization operator of graph signal processing. This method can select sensors while considering the localizations both in graph vertex domain and graph spectral domain and is fast, since eigendecomposition of graph Laplacian matrix is not required. We also propose an interpretation of the conventional node selection based on graph sampling theory by using the graph localization operators. Experiments on selected sensor location, execution time and prediction error comparisons are conducted to show the effectiveness of our approach.

Brain-computer interfacing (BCI) based on steady-state visual evoked potentials (SSVEPs) is one of the most promising techniques due to its high performance. A state-of-the-art is a BCI based on hybrid frequency and phase coded SSVEP, which needs a large set of calibration data as reference signals, so-called individual templates. The aim of this study is to propose an approach to calibration reduction by generating from individual templates corresponding to a part of commands (source templates) to new templates corresponding to the rest of commands. The new templates can be obtained by shifting the frequency and phase of the source template to the desired frequency and phase. In this way, time and cost for calibration can be greatly reduced. The experimental results suggested that the proposed approach successfully transferred the source template, closely achieving the performance using the full calibration dataset.

The estimation of covariance matrices is of prime importance to analyze the distribution of multivariate signals. In motor imagery-based brain-computer interfaces (MI-BCI), covariance matrices play a central role in the extraction of features from recorded electroencephalograms (EEGs); therefore, correctly estimating covariance is crucial for EEG classification. This letter discusses algorithms to average sample covariance matrices (SCMs) for the selection of the reference matrix in tangent space mapping (TSM)-based MI-BCI. Tangent space mapping is a powerful method of feature extraction and strongly depends on the selection of a reference covariance matrix. In general, the observed signals may include outliers; therefore, taking the geometric mean of SCMs as the reference matrix may not be the best choice. In order to deal with the effects of outliers, robust estimators have to be used. In particular, we discuss and test the use of geometric medians and trimmed averages (defined on the basis of several metrics) as robust estimators. The main idea behind trimmed averages is to eliminate data that exhibit the largest distance from the average covariance calculated on the basis of all available data. The results of the experiments show that while the geometric medians show little differences from conventional methods in terms of classification accuracy in the classification of electroencephalographic recordings, the trimmed averages show significant improvement for all subjects.

This paper presents a frequency recognition method of steady-state visual evoked potentials (SSVEPs) using binary subbands with canonical correlation analysis (CCA). The first subband contains all the target frequencies of SSVEPs. The second one includes the SSVEP signal corresponding to a desired number of higher order stimulus frequencies, which is obtained by filtering out of required range of lower order stimuli. The full dimension of artificial reference signals are used for first subband, whereas a reduced dimension of references is employed for second subband to compute canonical correlation. The weighted sum of the obtained correlation values are used to recognize the frequency of an SSVEP. The experimental results show the superiority of the proposed method compared to the state-of-the-art recognition methods.

We consider the problem of selecting optimal sensor placements. The proposed approach is based on the sampling theorem of graph signals. We choose sensors that maximize the graph cut-off frequency, i.e., the most informative sensors for predicting the values on unselected sensors. We study the existing methods in the context of graph signal processing and clarify the relationship between these methods and the proposed approach. The effectiveness of our approach is verified through numerical experiments, showing advantages in prediction error and execution time.

In this paper, inspired by sparse principal component analysis (SPCA) via the elastic net regularization, we propose a new criterion for sparsification of the kernel principal component analysis (KPCA) with the elastic net regularization that can simultaneously consider the data approximation and sparsification. We first show that KPCA also can be relaxed into a regression framework optimization problem, with a quadratic penalty; l1-norm can then be integrated into the regression criterion, leading to a new cost function. The minimization is iteratively conducted together with alternating direction method of multipliers. Experimental results for toy examples and real world data support the analysis.

One of the major challenges in kernel adaptive filtering is how to construct an efficient dictionary of observed input signals. In this paper, we propose novel dictionary adaptation rules for kernel adaptive filtering. The first algorithm can efficiently “move” elements of the dictionary to increase the approximation performance. The second algorithm mainly focuses on a nonstationary system, which can yield the increase of the dictionary size. The proposed method can eliminate unnecessary elements in the dictionary. Numerical examples support the efficacy of the proposed methods.

An asynchronous brain-computer interface (BCI) is one of the crucial challenges in biomedical signal processing. In asynchronous BCIs, a state when a user does not intend to input commands needs to be distinguished from a state when he/she does. These states are called non-control (NC) state and intentional control (IC) state respectively. In this paper, a new phase-based method to discriminate between IC/NC states for steady-state visual evoked potential (SSVEP) based asynchronous BCIs is proposed. The method has a two-step tree structure: in the first step, a SSVEP frequency is recognized with canonical correlation analysis (CCA), and in the next step, the state of a user is detected as IC or NC with a classifier such as SVM using phase information. The proposed method was tested on six healthy subjects and has been proved to be reliable in terms of sensitivity and specificity.

An adaptive online algorithm with a dictionary of observed signals for kernel principal subspace analysis is presented. A coefficient matrix for eigenfunctions is updated by a recursive least squares (RLS)-type algorithm and entries in the dictionary are adaptively added / removed preserving orthogonality of the eigenfunctions. It is shown that the orthogonalization can be implemented by analytically solvable (generalized) eigenvalues of 2×2 matrices, instead of the computation of the inverse squared root of matrix having the size of the dictionary. Numerical example is then illustrated to support the analysis.

Steady-state visual evoked potentials (SSVEPs) enable brain-computer interfaces to achieve efficient performance in command detection accuracy and information transfer rate (ITR). However, a limited bandwidth of SSVEPs causes a limited number of possible command in BCIs. Moreover since the amplitude of SSVEP at a particular frequency depends on users, some BCI commands could be executed easily (higher accuracy) and the others hardly (lower accuracy). To solve these problems, the present paper proposes a paradigm of BCI using phase-modulated visual stimuli. The commands of BCI are represented by a combination of binary codes. Two different phases of flickering stimuli correspond to binary codes. A command is determined by the sequence of binary codes which is demodulated from the phases of SSVEP. This protocol can be analogues of binary phase-shift keying (BPSK) in digital communications. Experimental results show that the proposed paradigm achieves even probabilities of command execution, although only a single electrode was used for SSVEP detection.

The present research work proposes a new fast fixed-point average-value learning algorithm on the compact Stiefel manifold based on a mixed retraction/lifting pair. Numerical comparisons between fixed-point algorithms based on the proposed non-associated retraction/lifting map pair and two associated retraction/lifting pairs confirm that the averaging algorithm based on a combination of mixed maps is remarkably less computationally demanding than the same averaging algorithm based on any of the constituent associated retraction/lifting pairs.

Analysis of electroencephalography (EEG) energy is a useful technique in the brain signal processing. In this paper, we present a novel data analysis method based on a dynamic multivariate empirical mode decomposition (D-MEMD) algorithm to analyze EEG energy of three different conscious states such as normal awake, comatose and brain death. By using D-MEMD, we can not only denoise the original EEG data but also calculate the EEG energy of subjects in a dynamic time series. Moreover, from the result, we distinguish three consciousness levels. The results of healthy subject in normal awake, comatose patient and brain death will be shown. The analyzed results illustrate the effectiveness and performance of the proposed method in calculation of EEG energy for evaluating consciousness level.

In this paper, we propose a Electroencephalography(EEG) signal processing method for the purpose of supporting the patient's EEG consciousness analysis. Approximate entropy(ApEn), as a complexity based method appears to have potential application to physiological and clinical time-series data. Therefore, we present an ApEn based statistical measure for patient's EEG consciousness analysis. However, it is found that high frequency noise such as electronic interference and its harmonic from the surrounding containing in the real-life recorded EEG lead to inconsistent ApEn result. To solve this problem, first we design a bandstop filter to filter high frequency noise. Then the proposed method is supported by analysis on a real world example of distinguishing between the brain consciousness states of coma and brain death. The experimental results demonstrate the effectiveness and performance of the proposed method in patient's EEG consciousness analysis.

The paper presents a novel concept implementing a phase locking value index estimation in application to brain-computer interfacing (BCI) motor imagery paradigm. We propose to decompose first the pairs of EEG channels using a bivariate empirical mode decomposition (BEMD) method. Next, the phase locking values (PLV) are estimated for the obtained intrinsic mode functions resulting in discriminating features drawn from EEG channel pairs representing the two different lateral hemispheres. Numerical results suggest that the PLV induced from BEMD can effectively detect phase synchrony between electrodes and is a promising feature for BCI implementation.

We present a novel method based on joint tensor diagonalization for selecting or weighting electroencephalogram (EEG) data to estimate the covariance matrices to accurately find common spatial pattern (CSP). CSP and its variants need a pair of covariance matrices of two different tasks, which are obtained as the average over trials. This trial average can affect the accurate estimation of covariance matrices and cause the decrease of classification accuracy in brain machine interfaces (BMIs) due to the non-stationarity of EEG or experimental environments. We focus on the fact that finding CSP is equivalent to joint diagonalization of a pair of covariance matrices, and extend it to joint diagonalization of data tensor at each trial to determine importance of each trial. Numerical experiment of motor imagery (MI) classification supports the proposed algorithm is effective.

We propose a regularization based on geometric structure for feature extraction in a sensor array for brain data recordings. The purpose of the study is to add a penalty term using distances between sensors as the geometric information for finding spatial weights. The regularization term is derived under the definition of neighbors of sensors. We evaluate the proposed regularization in common spatial pattern (CSP) which is a well-known feature extraction method for EEG based brain computer interface (BCI). We have demonstrated the CSP procedure with the regularization by simulation for artificial signals. The results show that the proposed method works better than standard CSP in extracting of a component generated in a certain brain spot. Moreover, the classification experimental results using dataset of motor imagery based BCI suggest that the proposed method achieved maximum improvement by 27% in the classification accuracy over the standard CSP in a setting of even when we use only five samples.

The aim of the present contribution is to extend the algorithm introduced in the paper S. Fiori and T. Tanaka, “An algorithm to compute averages on matrix Lie groups,” IEEE Transactions on Signal Processing, Vol. 57, No. 12, pp. 4734 - 4743, December 2009, to compute averages over the Stiefel manifold. The idea underlying the developed algorithms is that points on the Stiefel manifold are mapped onto a tangent space, where the average is taken, and then the average point on the tangent space is projected back to the Stiefel manifold. Based on this idea, a fixed-point algorithm is developed, and numerical examples are shown to support the analysis.

Measured electroencephalography (EEG) signals can be contaminated with other electrophysiological signal sources. This contamination decreases accuracy of neuroengineering applications such as brain computer interfaces. This paper focuses on the removal of electrooculography (EOG) that strongly appears in frontal electrodes EEG. To develop an EOG removal algorithm, we propose to utilize recently developed a multivariate extension of empirical mode decomposition (EMD) called MEMD. MEMD decomposes a multichannel signal into a set of intrinsic mode functions (IMF), and the number of IMFs is identical among the channels. We establish a criterion for choosing IMFs to separate an EOG-related component from the observed signal. Numerical examples confirm the proposed approach extracts EOG component better comparing to conventional blind source separation methods.

Adaptive online algorithms for simultaneously extracting nonlinear eigenvectors of kernel principal component analysis (KPCA) are developed. KPCA needs all the observed samples to represent basis functions, and the same scale of eigenvalue problem as the number of samples should be solved. This paper reformulates KPCA and deduces an expression in the Euclidean space, where an algorithm for tracking generalized eigenvectors is applicable. The developed algorithm here is least mean squares (LMS)-type and recursive least squares (RLS)-type. Numerical example is then illustrated to support the analysis.

The present research takes its moves from previous contributions by the present authors on two topics, namely, neural learning on differentiable manifolds by manifold retractions and averaging over differentiable manifolds. Learning on differentiable manifolds is a general theory that allows a neural system that insists on curved smooth spaces to adapt its parameters without violating the constraints on the geometry of the parameter spaces. In particular, the present contribution focuses on learning on the compact Stiefel manifold by manifold retraction with application to averaging `tall-skinny' matrices and generalizes some contributions recently appeared in the scientific literature about such a topic.

Brain-computer interfacing (BCI) is an emerging technology to translate non-invasively measured brain signals into commands, providing an additional communication channel. This paper develops a BCI platform to remotely control a mobile robot through the Internet. The platform consists of a server computer for EEG signal acquisition and signal processing/classification and a bluetooth controlled robot carrying an Android smartphone. The smartphone can give a visual feedback via the Internet to the user with its phone camera. The Android smartphone transfers the command from the server to the robot and transmits video stream to the server for the feedback to the user. It is also shown to employ the steady state visual evoked potentials (SSVEP) to classify commands. We propose a method for detecting the idle state when a user does not gaze any visual stimuli.

We propose a novel method for the classification of EEG signals during motor-imagery. For motor-imagery based brain computer interface (MI-BCI), a method called common spatial pattern (CSP), which finds spatial weights for electrodes, is effective, however CSP needs bandpass filtering as preprocessing. This paper addresses the problem to find parameters of the filter as well as the spatial weights. The filter is parameterize as weights for frequency spectra. Finding the optimal parameters is formulated as a constraint minimum variance problem. Then, the spatial and frequency weights are sought by alternately solving the generalized eigenvalue problem, and the cost function monotonically decreases by the alternative optimization. In our experiment of MI-BCI, the proposed method achieved maximum improvement by 6% in the classification accuracy over conventional methods.

An observed signal of the Doppler weather radar includes not only weather echoes but also a ground clutter. For accurate observation of weather data, we need to remove the effect of the ground clutter. In this paper, we propose to model the spectrum of an observed IQ signal as a mixture density function. To estimate the parameters of the density function, we apply the expectation-maximization (EM) algorithm in a maximum a posteriori (MAP) estimation with hyper parameters learned from the actual measurements of the ground clutter. Experimental results show that the proposed method works well in estimating the wind velocity, rainfall amount, and turbulence from the weather echo even when the spectrum of the weather echo is overlapped with that of the ground clutter in a lower frequency band.

Electroencephalography (EEG) is widely used in evaluating the absence of cerebral cortex function for the determination of brain death. Since EEG recorded signal is always corrupted by some artefacts and various interfering noise, extracting active or nonactive features from noisy EEG signals and evaluating their significance is therefore crucial in the process of brain death diagnosis. This article presents an EEG-based preliminary examination system associated with empirical mode decomposition (EMD) technique to extract informative brain activity features from real-world recorded clinical EEG data. Moreover, the power spectrum technique is applied to evaluate the significant differences between the group of comatose patients and the group of quasi-brain-deaths. Our experimental results show effectiveness and some promising directions of applying the EMD method to the clinical EEG analysis.

A novel method termed differential joint diagonalization (DJD) is introduced to find good initial values for joint diagonalization of time-varying correlation matrices. A key point of the method is finding a matrix which satisfies several conditions in the form of the matrix Riccati equation simultaneously. An alternate algorithm of differential joint diagonalization and joint diagonalization is proposed and shown by numerical simulation to outperform a standard joint diagonalization algorithm in tracking time-varying mixing models.

A problem of eye-movement muscular interference removal from EEG recordings is described. In many experiments in neuroscience it is crucial to separate different sources of electrical activity within human body in a situation when a very limited knowledge about nonlinear and nonstationary nature of the mixing process is available. A new two step extension to bivariate empirical mode decomposition is proposed to remove ocular artifacts from EEG with a use of fractional Gaussian noise as a reference first to preprocess EOG signal, which is next used in the second step as a reference to clean EEG signals. Results with EEG experimental data validate the proposed approach.

We propose a novel method for extracting a rhythmically oscillating signal from EEG recordings including multiple source signals which have similar frequencies. The main application of this method is brain computer interfaces (BCI), which use rhythmically oscillating signals such as alpha, mu, and beta waves, as feature signals. It is difficult to separate those components and/or extract an command-related component when these feature signals span the same frequency band. The main idea is to assume that signals generated in different brain parts have different phases, even though they have the same frequency. This hypothesis is effectively incorporated with the previously proposed rhythmic component extraction (RCE) method, which successfully extracts a signal oscillating at a certain frequency from multi-channel sensor signals in the BCI application. The signal model is firstly given and then this novel extraction method is formulated as an optimization problem. We apply the proposed method for the classification of multi-channel EEG signals between imaginary left/right hand movement. Our experiment suggests that the proposed method is effective in feature extraction for motor-imagery based BCI.

It is important to estimate the driver dozing practically. In the conventional study, the electroencephalogram (EEG) has been a promising indicator to driver dozing. Furthermore, it is known that frequency of the EEG is highly related to the sleep and the wake conditions. Therefore, we extract frequency components of the EEG from the whole cerebral cortex, by using the rhythmic component extraction (RCE), proposed by Tanaka et al. RCE extracts a component by combining multi-channel signals with weights that are optimally sought for such that the extracted component maximally contains the power in the frequency range of interest and suppresses that in unnecessary frequencies. As a result, we confirmed the interested frequency power is emphasized. These results imply that this method is available for analyzing the sleeping.

We propose a novel efficient method of blind signal extraction from multi-sensor networks when each observed signal consists of one global signal and local uncorrelated signals. Most of existing blind signal separation and extraction methods such as independent component analysis have constraints such as statistical independence, non-Gaussianity, and underdetermination, and they are not suitable for global signal extraction problem from noisy observations. We developed an estimation algorithm based on alternating iteration and the smart weighted averaging. The proposed method does not have strong assumptions such as independence or non-Gaussianity. Experimental results using a musical signal and a real electroencephalogram demonstrate the advantage of the proposed method.

A problem of information separation in multichannel recordings is important in engineering applications such as brain computer/machine interfaces (BCI/BMI). Whereas this problem is not entirely new, engineering approaches connecting the mental states of humans and the observed electroencephalography (EEG) recordings are still in their infancy, mostly due to problems with electrophysiological denoising. The electrophysiological signals captured in form of the EEG carry brain activity in form of the neurophysiological components which are usually embedded in much higher power electrical muscle activity components (electromyography - EMG; electrooculography - EOG; etc.). In this paper we present an approach to remove muscular interference caused by eye-movements from EEG recorded during auditory experiments in an eight channel recording setting. This is achieved by analyzing the correlation of the oscillatory modes within a multichannel signal in the Hilbert domain. Simulations in a real world auditory BCI setting support the analysis.

Rhythmic component extraction (RCE) is a method for extracting a signal oscillating at a certain frequency from multi-channel sensor signals. This method can be effectively used for detecting rhythmic signals such as alpha and beta waves, which are the feature signals in brain computer/machine interfaces (BCI/BMI). We are addressing a problem in developing an on-line adaptive algorithm for RCE. Since a rhythmic signal in the brain slowly varies, the signals extracted in adjacent frames should not largely different. We propose introducing a regularization term that evaluates the correlation between the signal extracted in the last step and the one to be extracted to achieve this. We show that the maximization of the cost function with the proposed regularization term is reduced to a generalized eigenvalue problem and experimental results from practical EEG data support this analysis.

In the present paper, we treat the problem of learning averages out of a set of symmetric positive-definite matrices (SPDMs). We discuss a possible learning technique based on the differential geometrical properties of the SPDM-manifold which was recently shown to possess a Lie-group structure under appropriate group definition. We first recall some relevant notions from differential geometry, mainly related to Lie-group theory, and then we propose a scheme of learning averages. Some numerical experiments will serve to illustrate the features of the learnt averages.

A committee of learning machines may be conceived of as a group of adaptive systems that adapt independently of each other and whose goal is to solve a common learning problem. Each machine in a committee computes a set of parameter-patterns belonging to a curved space. A natural question is how to combine the learnt patterns in order to obtain a better solution to the learning problem. In the present paper, we treat the case that the parameter space is the Lie group of unitary matrices. In order to combine the learnt patterns, we discuss a possible merging technique based on the differential geometrical structure of the parameter manifold.

In this experiment, we record the EEG data during the sleeping and the awakening for the drowsiness cognition. As is well known, analyzing the frequency components of the EEG is important for the sleeping. There are many studies to analyze the EEG frequency data for recognizing the sleeping quality, and so on. However, there is no established method for the analysis of the sleeping EEG. From these reason, we are going to apply the rhythmic component extraction (RCE), proposed by Tanaka et al., to extract the rhythmic component in the brain. RCE finds a component extracted by using a weighted sum of observed channel signals. The weights are optimized by maximizing the power in a certain frequency range of interest. In the experiment, the subjects lie back, relax in a chair, and sleep. We take the EEG recording at 30 channels. From the results, the EEG features, such as the alpha and the theta wave, are different between the sleeping and the waking. These rhythmic components are extracted by the Fourier transform and the RCE. By using the RCE, we confirmed the alpha wave and the theta wave when they are not extracted in a single channel signal. Furthermore, we confirmed the effective measurement locations by analyzing the RCE weights.

This letter develops a power method-based algorithm for tracking generalized eigenvectors when stochastic signals having unknown correlation matrices are observed. The proposed approach is based on the fact that the generalized eigenvalue problem is reduced to a standard eigenvalue problem, for which the power method can be easily applied by changing the metric of the vector space. The difficulty of applying the power method to this problem lies in the computational load of the inverse of the square root of a matrix at every update. The proposed algorithm can avoid the computation of eigenvalue decomposition to obtain the inverse of the matrix square root. Experimental results show that the proposed tracking algorithm gives a similar performance in tracking to the direct computation of singular value decomposition (SVD), while the computation order of the proposed algorithm is lower than the SVD.

A practical method for extracting and enhancing a rhythmic waveform appearing in multi-channel electroencephalogram (EEG) data is proposed. In order to facilitate clinical diagnosis and/or implement so-called brain computer interface (BCI), detecting the rhythmic activity from EEG data recorded in a noisy environment is crucial; however, classical signal processing techniques like linear filtering or the Fourier transform cannot detect such a rhythmic signal if the power of noise is so large. This paper presents a simple but practical method for extracting a rhythmic signal by fully exploiting the multi-variate nature of EEG data. The rhythmic component of interest is estimated as the weighted sum of multi-channel signals, and the optimal weights are then derived so as to maximize the power of the component. After the derivation is illustrated, adaptive weights, which give a new time-frequency analysis, are introduced. Moreover, the application to recently developed empirical mode decomposition (EMD) is presented. Experimental results on real EEG data support the analysis.

The present paper aims at introducing a novel procedure for designing an averaging algorithm for a committee of learning machines under the assumption that the machines share a common parameter-space, namely, the group of special orthogonal matrices SO(p). The averaging procedure inputs the patterns learnt by the machines in the committee and outputs a mean-matrix that represents an average of machines’ patterns. Since the space SO(p) is a curved manifold, averaging does not carry on the usual (Euclidean) meaning and should be designed on the basis of the parameter-space’s geometric properties.

A new method to extend the empirical mode decomposition (EMD) into the complex domain is proposed. Unlike the existing method for EMD in the complex domain, this is achieved in a generic way so that the mathematical development of this method mirrors the algorithm defined for EMD in the real domain. The so derived intrinsic mode functions (IMFs) are complex by design and are shown to provide a consistent framework for handling both real and complex data. The simulations on real world complex-valued signals illustrate the applications of the technique.

The theory and derivation of a novel method for approximate joint diagonalization (AJD) on the orthogonal group of matrices are presented. The proposed algorithms are fast and simple, hence, easy to implement. We introduce a least-squares-type cost function, which is to be minimized under the constraint that the matrix to be sought for is orthogonal. A gradient flow for optimizing such cost function is derived and its stability is analyzed within the framework of differential geometry. It is proposed to numerically approximate the gradient flow by using a geodesic-based and an Euler-like update algorithms. Numerical examples about blind source separation of speech signals are illustrated to support the analysis.

We consider the problem of recovering a high-resolution (HR) frame from a sequence of low-resolution (LR) frames. It is challenging to design a super-resolution (SR) algorithm for arbitrary video sequences. Video frames in general cannot be related through global parametric transformation due to the arbitrary individual pixel movement between frame pairs. Hence a local motion model needs to be used for frame alignment. An accurate alignment is the key to success of reconstruction-based super-resolution algorithms. Motivated by this challenge we propose to employ region-matching technique for image registration in this paper. The proposed algorithm consists of the alignment step to produce a blurred version of the HR frame and the restoration step to estimate the HR frame. The experimental results of the proposed algorithm are compared with the results of using affine, block matching, and optical flow motion models. It is shown that the use of region matching for SR is very promising in producing higher quality images.

A method for the empirical mode decomposition (EMD) of complex-valued data is proposed. This is achieved based on the filter bank interpretation of the EMD mapping and by making use of the relationship between the positive and negative frequency component of the Fourier spectrum. The so-generated intrinsic mode functions (IMFs) are complex-valued, which facilitates the extension of the standard EMD to the complex domain. The analysis is supported by simulations on both synthetic and real-world complex-valued signals

A new on-line learning algorithm that yields a reduced-rank Wiener filter (RRWF) is proposed. The RRWF is defined as the matrix of prescribed rank that provides the best least-squares approximation of a given signal. This implies that an RRWF determines only a subspace, but is not endowed with information of basis functions or axes for the subspace. In other words, even if we want to reduce the rank of the estimated RRWF, we should learn another RRWF of "more reduced rank" again. Our goal in this paper is therefore to establish a learning rule that simultaneously tracks basis functions yielding a matrix that gives an RRWF. To this end, we reformulate the optimization problem of RRWFs, which will be solved by a gradient-based algorithm derived within the framework of differential geometry. Numerical examples are illustrated to support the proposals in the paper

Many interesting signals share the property of being sparsely active. The search for such sparse components within a data set commonly involves a linear or nonlinear projection step in order to fulfill the sparseness constraints. In addition to the proximity measure used for the projection, the result of course is also intimately connected with the actual definition of the sparseness criterion. In this work, we introduce a novel sparseness measure and apply it to the problem of finding a sparse projection of a given signal. Here, sparseness is defined as the fixed ratio of p-over 2-norm, and existence and uniqueness of the projection holds. This framework extends previous work by Hoyer in the case of p = 1, where it is easy to give a deterministic, more or less closed-form solution. This is not possible for p ≠ 1, so we introduce an algorithm based on alternating projections onto spheres (POSH), which is similar to the projection onto convex sets (OCS). Although the assumption of convexity does not hold in our setting, we observe not only convergence of the algorithm, but also convergence to the correct minimal distance solution. Indications for a proof of this surprising property are given. Simulations confirm these results.

We present a novel lossless compression algorithm which compresses sequences of three-dimensional (3D) volumes collected during a functional magnetic resonance imaging (fMRI) experiment. The large data sets involved in this popular biomedical application necessitate fast and efficient compression methods. We propose to use 3D prediction, temporal decorrelation and entropy coding with context modeling for encoding the fMRI scans after preprocessing with the region-of-interest (ROI) masking. The proposed algorithm is conceptually simple and can achieve fast implementation and efficient coding performance. We illustrate computer simulations to show advantages over conventional coding methods.

A class of lapped transforms for image coding, which are characterized by variable-length synthesis filters, is introduced. In this class, the synthesis filter bank (FB) is first defined with an arbitrary combination of finite impulse response synthesis filters of perfect reconstruction FBs. An analysis FB is then obtained using direct matrix inversion or iterative implementation of Neumann series expansion. Moreover, to improve compression, we introduce a unitary transform that follows the analysis FB. This class enables a greater freedom of design than previously presented variable-length lapped transforms. We illustrate several design examples and present experimental results for image coding, which indicate that the proposed transforms are promising and comparable with conventional subband transforms including wavelets.

A novel stereophonic noise reduction method is proposed. This method is based upon a combination of a subspace approach realized in a sliding window operation and two-channel adaptive signal enhancing. The signal obtained from the signal subspace is used as the input signal to the adaptive signal enhancer for each channel, instead of noise, as in the ordinary adaptive noise canceling scheme. Simulation results based upon real stereophonic speech contaminated by noise components show that the proposed method gives improved enhancement quality in terms of both segmental gain and cepstral distance performance indices in comparison with conventional nonlinear spectral subtraction approaches.

We introduce a new extended model for independent component analysis (ICA) and/or blind source separation (BSS), in which the assumption of the standard ICA model that the source signals are mutually independent (or spatio-temporally uncorrelated) is relaxed. The source is presumed to be the sum of some independent and/or dependent subcomponents. We show a practical solution for this class of blind separation problem by using subband decomposition (SD) and the independence test by analyzing global mixing-demixing matrices obtained for various subbands or multi-bands. This is a very simple but efficient technique, and users just apply the proposed method to conventional ICA/BSS algorithms as pre- and post-processing. The proposed method has been tested for blind separation problems with partially dependent sources. The results indicate that the method is promising for the signal separation problem of speech, image, EEG data, etc.