Steady-State Visual Evoked Potential Classification Using Complex Valued Convolutional Neural Networks.

The steady-state visual evoked potential (SSVEP), which is a kind of event-related potential in electroencephalograms (EEGs), has been applied to brain-computer interfaces (BCIs). SSVEP-based BCIs currently perform the best in terms of information transfer rate (ITR) among various BCI implementation methods. Canonical component analysis (CCA) or spectrum estimation, such as the Fourier transform, and their extensions have been used to extract features of SSVEPs. However, these signal extraction methods have a limitation in the available stimulation frequency; thus, the number of commands is limited. In this paper, we propose a complex valued convolutional neural network (CVCNN) to overcome the limitation of SSVEP-based BCIs. The experimental results demonstrate that the proposed method overcomes the limitation of the stimulation frequency, and it outperforms conventional SSVEP feature extraction methods.

Neural decoding of code modulated visual evoked potentials by spatio-temporal inverse filtering for brain computer interfaces.

This study addresses neural decoding of a code modulated visual evoked potentials (c-VEPs). c-VEP was recently developed, and applied to brain computer interfaces (BCIs). c-VEP BCI exhibits faster communication speed than existing VEP-based BCIs. In c-VEP BCI, the canonical correlation analysis (CCA) that maximizes the correlation between an averaged signal and single trial signals is often used for the spatial filter. However, CCA does not utilize information of given PN sequence, and hence, the filtered signal may not have properties of PN sequence. In this paper, we propose a decoding method to restore the given PN sequence from the observed VEP. We compare linear and nonlinear spatio-temporal inverse filtering methods. For the linear method, the least mean square error and lasso are used to obtain the filter coefficients. For the non-linear method, the artificial neural network is used. The proposed methods exhibited better decoding performance, and higher classification accuracies than conventional CCA spatial filtered c-VEP BCI.

Reliability-based automatic repeat request for short code modulation visual evoked potentials in brain computer interfaces.

We propose two methods to improve code modulation visual evoked potential brain computer interfaces (cVEP BCIs). Most of BCIs average brain signals from several trials in order to improve the classification performance. The number of averaging defines the trade-off between input speed and accuracy, and the optimal averaging number depends on individual, signal acquisition system, and so forth. Firstly, we propose a novel dynamic method to estimate the averaging number for cVEP BCIs. The proposed method is based on the automatic repeat request (ARQ) that is used in communication systems. The existing cVEP BCIs employ rather longer code, such as 63-bit M-sequence. The code length also defines the trade-off between input speed and accuracy. Since the reliability of the proposed BCI can be controlled by the proposed ARQ method, we introduce shorter codes, 32-bit M-sequence and the Kasami-sequence. Thanks to combine the dynamic averaging number estimation method and the shorter codes, the proposed system exhibited higher information transfer rate compared to existing cVEP BCIs.

Mahalanobis distance on extended Grassmann manifolds for variational pattern analysis.

In pattern classification problems, pattern variations are often modeled as a linear manifold or a low-dimensional subspace. Conventional methods use such models and define a measure of similarity or dissimilarity. However, these similarity measures are deterministic and do not take into account the distribution of linear manifolds or low-dimensional subspaces. Therefore, if the distribution is not isotopic, the distance measurements are not reliable, as well as vector-based distance measurement in the Euclidean space. We previously systematized the representations of variational patterns using the Grassmann manifold and introduce the Mahalanobis distance to the Grassmann manifold as a natural extension of Euclidean case. In this paper, we present two methods that flexibly extend the Mahalanobis distance on the extended Grassmann manifolds. These methods can be used to measure pattern (dis)similarity on the basis of the pattern structure. Experimental evaluation of the performance of the proposed methods demonstrated that they exhibit a lower error classification rate.

A novel EEG-based spelling system using N100 and P300.

P300-speller is one of the most popular EEG-based spelling systems proposed by Farwell and Donchin. P300-speller has a 6 × 6 character matrix and requires at least 12 flashes to input one character. This restricts increasing of the information transfer rate (ITR) by decreasing the number of flashes. In this paper, a novel spelling system is proposed to reduce the number of flashes by sequential stimulation of images. In order to determine the command, the proposed system utilizes two kinds of the event-related brain potentials (ERP), N100 and P300 whereas P300-speller utilizes only P300. Thanks to using both N100 and P300, the proposed system achieves higher accuracy and faster spelling speed than P300-speller. Our experiment by ten subjects showed that ITR of the proposed system is an average of 0.25 bit/sec improved compared to P300-speller.

Adaptive subset kernel principal component analysis for time-varying patterns.

Kernel principal component analysis (KPCA) and its online learning algorithms have been proposed and widely used. Since KPCA uses training samples for bases of the operator, its online learning algorithms require the preparation of all training samples beforehand. Subset KPCA (SubKPCA), which uses a subset of samples for the basis set, has been proposed and has demonstrated better performance with less computational complexity. In this paper, we extend SubKPCA to an online version and propose methods to add and exchange a sample in the basis set. Since the proposed method uses the basis set, we do not need to prepare all training samples beforehand. Therefore, the proposed method can be applied to time-varying patterns, in contrast to existing online KPCA algorithms. Experimental results demonstrate the advantages of the proposed method.

EEG auditory steady state responses classification for the novel BCI.

An auditory modality brain computer interface (BCI) is a novel and interesting paradigm in neurotechnology applications. The paper presents a concept of auditory steady state responses (ASSR) utilization for the novel BCI paradigm. Two EEG feature extraction approaches based on a bandpass filtering and an AR spectrum estimation are tested together with two classification schemes in order to validate the proposed auditory BCI paradigm. The resulting good classification scores of users intentional choices, of attending or not to the presented stimuli, support the hypothesis of the ASSR stimuli validity for a solid BCI paradigm.

Kernel Wiener filter and its application to pattern recognition.

The Wiener filter (WF) is widely used for inverse problems. From an observed signal, it provides the best estimated signal with respect to the squared error averaged over the original and the observed signals among linear operators. The kernel WF (KWF), extended directly from WF, has a problem that an additive noise has to be handled by samples. Since the computational complexity of kernel methods depends on the number of samples, a huge computational cost is necessary for the case. By using the first-order approximation of kernel functions, we realize KWF that can handle such a noise not by samples but as a random variable. We also propose the error estimation method for kernel filters by using the approximations. In order to show the advantages of the proposed methods, we conducted the experiments to denoise images and estimate errors. We also apply KWF to classification since KWF can provide an approximated result of the maximum a posteriori classifier that provides the best recognition accuracy. The noise term in the criterion can be used for the classification in the presence of noise or a new regularization to suppress changes in the input space, whereas the ordinary regularization for the kernel method suppresses changes in the feature space. In order to show the advantages of the proposed methods, we conducted experiments of binary and multiclass classifications and classification in the presence of noise.

Blind extraction of global signal from multi-channel noisy observations.

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.

Feature extraction using constrained approximation and suppression.

In this paper, we systematize a family of constrained quadratic classifiers that belong to the class of one-class classifiers. One-class classifiers such as the single-class support vector machine or the subspace methods are widely used for pattern classification and detection problems because they have many advantages over binary classifiers. We interpret subspace methods as rank-constrained quadratic classifiers in the framework. We also introduce two constraints and a method of suppressing the effect of competing classes to make them more accurate and retain their advantages over binary classifiers. Experimental results demonstrate the advantages of our methods over conventional classifiers.

Extraction of steady state visually evoked potential signal and estimation of distribution map from EEG data.

We propose a signal extraction method from multi-channel EEG signals and apply to extract Steady State Visually Evoked Potential (SSVEP) signal. SSVEP is a response to visual stimuli presented in the form of flushing patterns. By using several flushing patterns with different frequency, brain machine (computer) interface (BMI/BCI) can be realized. Therefore it is important to extract SSVEP signals from multi-channel EEG signals. At first, we estimate the power of the objective signal in each electrode. Estimation of the power is helpful in not only extraction of the signal but also drawing a distribution map of the signal, finding electrodes which have large SNR, and ranking electrodes in sort of information with respect to the power of the signal. Experimental results show that the proposed method 1) estimates more accurate power than existing methods, 2) estimates the global signal which has larger SNR than existing methods, and 3) allows us to draw a distribution map of the signal, and it conforms the biological theory.

Simultaneous pattern classification and multidomain association using self-structuring kernel memory networks.

In this paper, a novel exemplar-based constructive approach using kernels is proposed for simultaneous pattern classification and multidomain pattern association tasks. The kernel networks are constructed on a modular basis by a simple one-shot self-structuring algorithm motivated from the traditional Hebbian principle and then, they act as the flexible memory capable of generalization for the respective classes. In the self-structuring kernel memory (SSKM), any arduous and iterative network parameter tuning is not involved for establishing the weight connections during the construction, unlike conventional approaches, and thereby, it is considered that the networks do not inherently suffer from the associated numerical instability. Then, the approach is extended for multidomain pattern association, in which a particular domain input cannot only activate some kernel units (KUs) but also the kernels in other domain(s) via the cross-domain connection(s) in between. Thereby, the SSKM can be regarded as a simultaneous pattern classifier and associator. In the simulation study for pattern classification, it is justified that an SSKM consisting of distinct kernel networks can yield relatively compact-sized pattern classifiers, while preserving a reasonably high generalization capability, in comparison with the approach using support vector machines (SVMs).