Motor Imagery Dataset Research Articles

Lightweight Source-Free Transfer for Privacy-Preserving Motor Imagery Classification

Transfer learning utilizes data or knowledge in one problem to help solve a related problem. It is particularly useful in electroencephalogram (EEG) based motor imagery (MI) classification, to handle high intra-subject and/or cross-subject variations. This paper considers offline unsupervised cross-subject MI classification, i.e., we have labeled EEG trials from several source subjects, but only unlabeled EEG trials from the target subject. Existing transfer learning approaches usually make use of the source domain data directly in target model learning. To protect the privacy of the source subjects, we propose lightweight source-free transfer (LSFT), which first generates source models locally and encapsulates them as model application programming interfaces (APIs), then constructs a virtual intermediate domain to transfer the knowledge in the source domains to the target domain, and finally performs feature adaptation learning. Compared with existing deep transfer learning approaches, LSFT does not need to transfer from massive source data or models, is computationally efficient, and has a small number of parameters. Experiments on four benchmark MI datasets demonstrated that LSFT outperformed 13 different approaches, including several state-of-the-art transfer learning approaches that make use of the source domain samples or model parameters directly.

Motor Imagery Classification Based on EEG Sensing with Visual and Vibrotactile Guidance.

Motor imagery (MI) is a technique of imagining the performance of a motor task without actually using the muscles. When employed in a brain-computer interface (BCI) supported by electroencephalographic (EEG) sensors, it can be used as a successful method of human-computer interaction. In this paper, the performance of six different classifiers, namely linear discriminant analysis (LDA), support vector machine (SVM), random forest (RF), and three classifiers from the family of convolutional neural networks (CNN), is evaluated using EEG MI datasets. The study investigates the effectiveness of these classifiers on MI, guided by a static visual cue, dynamic visual guidance, and a combination of dynamic visual and vibrotactile (somatosensory) guidance. The effect of filtering passband during data preprocessing was also investigated. The results show that the ResNet-based CNN significantly outperforms the competing classifiers on both vibrotactile and visually guided data when detecting different directions of MI. Preprocessing the data using low-frequency signal features proves to be a better solution to achieve higher classification accuracy. It has also been shown that vibrotactile guidance has a significant impact on classification accuracy, with the associated improvement particularly evident for architecturally simpler classifiers. These findings have important implications for the development of EEG-based BCIs, as they provide valuable insight into the suitability of different classifiers for different contexts of use.

UNCER: A framework for uncertainty estimation and reduction in neural decoding of EEG signals

EEG decoding systems based on deep neural networks have been widely used in the decision-making of brain-computer interfaces (BCI). Their predictions, however, can be unreliable given the significant variance and noise in EEG signals. Previous works on EEG analysis mainly focus on the exploration of noise patterns in the source signal, while the uncertainty during the decoding process is largely unexplored. Automatically detecting and reducing such decoding uncertainty is important for BCI motor imagery applications such as robotic arm control etc. In this work, we proposed an uncertainty estimation and reduction model (UNCER) to quantify and mitigate the uncertainty during the EEG decoding process. It utilized a combination of dropout oriented method and Bayesian neural network for uncertainty estimation to incorporate both the uncertainty in the input signal and the uncertainty in the model parameters. We further proposed an adaptive data augmentation-based approach for uncertainty reduction. The model can be integrated into the current widely used EEG neural decoders without change of the architecture. We performed extensive experiments for uncertainty estimation and its reduction in both intra-subject EEG decoding and cross-subject EEG decoding on two public motor imagery datasets, where the proposed model achieves significant improvement both in the quality of estimated uncertainty and the effectiveness of uncertainty reduction.

E‐CNNet: Time‐reassigned Multisynchrosqueezing transform‐based deep learning framework for MI‐BCI task classification

AbstractThe classification of electroencephalograms‐based motor imagery signals poses a significant issue in the design and development of brain‐computer interfaces. Neural Networks are observed to be successful in classification of motor imagery brain signals. However, the existing motor imagery data sets are of limited size and suffer from low signal to noise ratio. Hence, to achieve high performance with small datasets, this paper proposes a novel combination of time frequency analysis along with deep learning network to perform the brain signal classification task. The proposed framework consists of two parts: (1) Time‐reassigned Multisynchrosqueezing Transform to efficiently capture the dynamic properties of non‐stationary EEG; and (2) A new hybrid model E‐CNNet is proposed for feature extraction and classification of brain signals. A robust classification pipeline is constructed to classify the extracted features into respective motor imagery tasks. Ensemble of boosting classifiers is used to generate the final predictions, thus, reducing the variance and bias of individual classifiers. The performance of the proposed methodology is evaluated using two publicly available datasets: BCI competition III, dataset IIIa and BCI competition IV, dataset IIa. We have obtained the average classification accuracy of 94.44% on BCI competition III, dataset IIIa and 89.25% on BCI competition IV, dataset IIa.

FBMSNet: A Filter-Bank Multi-Scale Convolutional Neural Network for EEG-Based Motor Imagery Decoding.

Motor imagery (MI) is a mental process widely utilized as the experimental paradigm for brain-computer interfaces (BCIs) across a broad range of basic science and clinical studies. However, decoding intentions from MI remains challenging due to the inherent complexity of brain patterns relative to the small sample size available for machine learning. This paper proposes an end-to-end Filter-Bank Multiscale Convolutional Neural Network (FBMSNet) for MI classification. A filter bank is first employed to derive a multiview spectral representation of the EEG data. Mixed depthwise convolution is then applied to extract temporal features at multiple scales, followed by spatial filtering to mitigate volume conduction. Finally, with the joint supervision of cross-entropy and center loss, FBMSNet obtains features that maximize interclass dispersion and intraclass compactness. We compare FBMSNet with several state-of-the-art EEG decoding methods on two MI datasets: the BCI Competition IV 2a dataset and the OpenBMI dataset. FBMSNet significantly outperforms the benchmark methods by achieving 79.17% and 70.05% for four-class and two-class hold-out classification accuracy, respectively. These results demonstrate the efficacy of FBMSNet in improving EEG decoding performance toward more robust BCI applications. The FBMSNet source code is available at https://github.com/Want2Vanish/FBMSNet.

Intra- and inter-subject common spatial pattern for reducing calibration effort in MI-based BCI.

One major problem limiting the practicality of a brain-computer interface (BCI) is the need for large amount of labeled data to calibrate its classification model. Although the effectiveness of transfer learning (TL) for conquering this problem has been evidenced by many studies, a highly recognized approach has not yet been established. In this paper, we propose a Euclidean alignment (EA)-based Intra- and inter-subject common spatial pattern (EA-IISCSP) algorithm for estimating four spatial filters, which aim at exploiting Intra- and inter-subject similarities and variability to enhance the robustness of feature signals. Based on the algorithm, a TL-based classification framework was developed for enhancing the performance of motor imagery (MI) BCIs, in which the feature vector extracted by each filter is dimensionally reduced by linear discriminant analysis (LDA) and a support vector machine (SVM) is used for classification. The performance of the proposed algorithm was evaluated on two MI data sets and compared with that of three state-of-the-art TL algorithms. Experimental results showed that the proposed algorithm significantly outperforms these competing algorithms for training trials per class from 15 to 50 and can reduce the amount of training data while maintaining an acceptable accuracy, thus facilitating the practical application of MI-based BCIs.

Hybrid EEG-fNIRS Brain Computer Interface Based on Common Spatial Pattern by Using EEG-Informed General Linear Model

Hybrid brain computer interfaces (BCI) utilizing the high temporal resolution of electroencephalography (EEG) and the high spatial resolution of near-infrared spectroscopy (fNIRS) are preferred over single-modal BCIs. However, due to the large dimensionality of the multi-class statistical features commonly used in fNIRS signals, it is easy to cause overfitting of the EEG-fNIRS hybrid BCI classifier. Therefore, a low-dimensional feature extraction method for fNIRS based on the EEG-informed fNIRS general linear model (GLM) analysis is proposed in this paper. First, a regression coefficient matrix is obtained by using the EEG-informed fNIRS GLM with a time window added, and the common spatial pattern (CSP) features of this regression coefficient matrix are extracted as the fNIRS features. Lastly, the fNIRS features were combined with the CSP features extracted from the optimal narrow band of EEG as hybrid features, and the support vector machine (SVM) method is used to classify the samples with hybrid features. The proposed method was tested on a publicly available motor imagery dataset. The classification accuracy using fNIRS signals alone reached 68.79% (oxygenated hemoglobin) and 68.62% (deoxygenated hemoglobin), and the classification accuracy of combining EEG-fNIRS features reached 79.48%, which was higher than other existing methods using the same dataset. By using this fNIRS feature extraction method, the problem of poor performance of CSP on fNIRS signals is solved, which not only enriches the processing methods of fNIRS signals, but also improves the classification accuracy of hybrid EEG-fNIRS BCI in motor imagery tasks.

Enhanced Motor Imagery Decoding by Calibration Model Assisted with Tactile ERD.

In this study, we propose a tactile-assisted calibration method for a motor imagery (MI) based Brain-Computer Interface (BCI) system. In the proposed calibration, tactile stimulation was applied to the hand wrist to assist the subjects in the MI task, which is named SA-MI task. Then, classifier training in the SA-MI Calibration was performed using the SA-MI data, while the Conventional Calibration employed the MI data. After the classifiers were trained, the performance was evaluated on a common MI dataset. Our study demonstrated that the SA-MI Calibration significantly improved the performance as compared with the Conventional Calibration, with a decoding accuracy of (78.3% vs. 71.3%). Moreover, the average calibration time could be reduced by 40%. This benefit of the SA-MI Calibration effect was further validated by an independent control group, which showed no improvement when tactile stimulation was not applied during the calibration phase. Further analysis showed that when compared with MI, greater motor-related cortical activation and higher R2 value in the alpha-beta frequency band were induced in SA-MI. Indeed, the SA-MI Calibration could significantly improve the performance and reduce the calibration time as compared with the Conventional Calibration. The proposed tactile stimulation-assisted MI Calibration method holds great potential for a faster and more accurate system setup at the beginning of BCI usage.

A Multi-Source Transfer Joint Matching Method for Inter-Subject Motor Imagery Decoding.

Individual differences among different subjects pose a great challenge to motor imagery (MI) decoding. Multi-source transfer learning (MSTL) is one of the most promising ways to reduce individual differences, which can utilize rich information and align the data distribution among different subjects. However, most MSTL methods in MI-BCI combine all data in the source subjects into a single mixed domain, which will ignore the effect of important samples and the large differences in multiple source subjects. To address these issues, we introduce transfer joint matching and improve it to multi-source transfer joint matching (MSTJM) and weighted MSTJM (wMSTJM). Different from previous MSTL methods in MI, our methods align the data distribution for each pair of subjects, and then integrate the results by decision fusion. Besides that, we design an inter-subject MI decoding framework to verify the effectiveness of these two MSTL algorithms. It mainly consists of three modules: covariance matrix centroid alignment in the Riemannian space, source selection in the Euclidean space after tangent space mapping to reduce negative transfer and computation overhead, and further distribution alignment by MSTJM or wMSTJM. The superiority of this framework is verified on two common public MI datasets from BCI competition IV. The average classification accuracy of the MSTJM and wMSTJ methods outperformed other state-of-the-art methods by at least 4.24% and 2.62% respectively. It's promising to advance the practical applications of MI-BCI.

Improving Generalization of CNN-Based Motor-Imagery EEG Decoders via Dynamic Convolutions.

Deep Convolutional Neural Networks (CNNs) have recently demonstrated impressive results in electroencephalogram (EEG) decoding for several Brain-Computer Interface (BCI) paradigms, including Motor-Imagery (MI). However, neurophysiological processes underpinning EEG signals vary across subjects causing covariate shifts in data distributions and hence hindering the generalization of deep models across subjects. In this paper, we aim to address the challenge of inter-subject variability in MI. To this end, we employ causal reasoning to characterize all possible distribution shifts in the MI task and propose a dynamic convolution framework to account for shifts caused by the inter-subject variability. Using publicly available MI datasets, we demonstrate improved generalization performance (up to 5%) across subjects in various MI tasks for four well-established deep architectures.

Speech2EEG: Leveraging Pretrained Speech Model for EEG Signal Recognition.

Identifying meaningful brain activities is critical in brain-computer interface (BCI) applications. Recently, an increasing number of neural network approaches have been proposed to recognize EEG signals. However, these approaches depend heavily on using complex network structures to improve the performance of EEG recognition and suffer from the deficit of training data. Inspired by the waveform characteristics and processing methods shared between EEG and speech signals, we propose Speech2EEG, a novel EEG recognition method that leverages pretrained speech features to improve the accuracy of EEG recognition. Specifically, a pretrained speech processing model is adapted to the EEG domain to extract multichannel temporal embeddings. Then, several aggregation methods, including the weighted average, channelwise aggregation, and channel-and-depthwise aggregation, are implemented to exploit and integrate the multichannel temporal embeddings. Finally, a classification network is used to predict EEG categories based on the integrated features. Our work is the first to explore the use of pretrained speech models for EEG signal analysis as well as the effective ways to integrate the multichannel temporal embeddings from the EEG signal. Extensive experimental results suggest that the proposed Speech2EEG method achieves state-of-the-art performance on two challenging motor imagery (MI) datasets, the BCI IV-2a and BCI IV-2b datasets, with accuracies of 89.5% and 84.07% , respectively. Visualization analysis of the multichannel temporal embeddings show that the Speech2EEG architecture can capture useful patterns related to MI categories, which can provide a novel solution for subsequent research under the constraints of a limited dataset scale.

MDTL: A Novel and Model-Agnostic Transfer Learning Strategy for Cross-Subject Motor Imagery BCI.

In recent years, deep neural network-based transfer learning (TL) has shown outstanding performance in EEG-based motor imagery (MI) brain-computer interface (BCI). However, due to the long preparation for pre-trained models and the arbitrariness of source domain selection, using deep transfer learning on different datasets and models is still challenging. In this paper, we proposed a multi-direction transfer learning (MDTL) strategy for cross-subject MI EEG-based BCI. This strategy utilizes data from multi-source domains to the target domain as well as from one multi-source domain to another multi-source domain. This strategy is model-independent so that it can be quickly deployed on existing models. Three generic deep learning models for MI classification (DeepConvNet, ShallowConvNet, and EEGNet) and two public motor imagery datasets (BCIC IV dataset 2a and Lee2019) are used in this study to verify the proposed strategy. For the four-classes dataset BCIC IV dataset 2a, the proposed MDTL achieves 80.86%, 81.95%, and 75.00% mean prediction accuracy using the three models, which outperforms those without MDTL by 5.79%, 6.64%, and 11.42%. For the binary-classes dataset Lee2019, MDTL achieves 88.2% mean accuracy using the model DeepConvNet. It outperforms the accuracy without MDTL by 23.48%. The achieved 81.95% and 88.2% are also better than the existing deep transfer learning strategy. Besides, the training time of MDTL is reduced by 93.94%. MDTL is an easy-to-deploy, scalable and reliable transfer learning strategy for existing deep learning models, which significantly improves model performance and reduces preparation time without changing model architecture.

Deep Common Spatial Pattern based Motor Imagery Classification with Improved Objective Function

Survey/review study Deep Common Spatial Pattern Based Motor Imagery Classification with Improved Objective Function Nanxi Yu 1,2, Rui Yang 1, and Mengjie Huang 1,* 1 School of Electrical Engineering, Electronics & Computer Science, University of Liverpool, Liverpool, L69 3BX, United Kingdom 2 Department of Biostatistics, Graduate School of Arts and Sciences, Yale University, New Haven, CT 06511, United States * Correspondence: Mengjie.Huang@liverpool.ac.uk Received: 12 October 2022 Accepted: 28 November 2022 Published: 22 December 2022 Abstract: Common spatial pattern (CSP) technique has been very popular in terms of electroencephalogram (EEG) features extraction in motor imagery (MI)-based brain-computer interface (BCI). Through the simultaneous diagonalization of the covariance matrices, CSP intends to transform data into another mapping with data of different categories having maximal differences in their measures of dispersion. This paper shows the objective function realized by original CSP method could be inaccurate by regularizing the estimated spatial covariance matrix from EEG data by trace, leading to some flaws in the features to be extracted. In order to deal with this problem, a novel deep CSP (DCSP) model with optimal objective function is proposed in this paper. The benefits of the proposed DCSP method over original CSP method are verified with experiments on two EEG based MI datasets where the classification accuracy is effectively improved.

Cognitive Computing for Brain–Computer Interface-Based Computational Social Digital Twins Systems

To accurately and effectively analyze electroencephalogram (EEG) with high complexity, large amount of data, and strong uncertainty, brain–computer interface (BCI) cognitive computing and its signal analysis algorithms are studied based on the digital twins (DTs) cognitive computing platform. To avoid the influence of noise on EEG analysis results, it is necessary to use filtering and defalsification methods to process EEG. Four methods, including Butterworth filter, finite impulse response (FIR) filter, elliptic filter, and wavelet decomposition, are summarized. Based on the Riemann manifold theory, a feature extraction algorithm under transfer learning based on tangent space selection (TL-TSS) is proposed. In the process of decoding EEG, an EEG decoding method combining entropy measure and singular spectrum analysis (SSA) is proposed. An algorithm performance is tested on the motor imagery dataset of the two International BCI Competitions. It is found that when the training sample size accounts for 5%, the TL-TSS algorithm proposed in this work is superior to other algorithms in classification accuracy. In particular, compared with common spatial pattern (CSP) algorithm, it has great advantages. The classification accuracy of A2, A4, A8, and A9 users is the best, and especially for A8 users, the classification accuracy reaches 97.88%. In summary, in the EEG interface technology of DT cognitive computing platform, the combination of cognitive computing and deep learning can improve the recognition and analysis effect of EEG, which is of great value for further optimization of DT cognitive computing system.

Subject adaptation convolutional neural network for EEG-based motor imagery classification

Objective. Deep transfer learning has been widely used to address the nonstationarity of electroencephalogram (EEG) data during motor imagery (MI) classification. However, previous deep learning approaches suffer from limited classification accuracy because the temporal and spatial features cannot be effectively extracted. Approach. Here, we propose a novel end-to-end deep subject adaptation convolutional neural network (SACNN) to handle the problem of EEG-based MI classification. Our proposed model jointly optimizes three modules, i.e. a feature extractor, a classifier, and a subject adapter. Specifically, the feature extractor simultaneously extracts the temporal and spatial features from the raw EEG data using a parallel multiscale convolution network. In addition, we design a subject adapter to reduce the feature distribution shift between the source and target subjects by using the maximum mean discrepancy. By minimizing the classification loss and the distribution discrepancy, the model is able to extract the temporal-spatial features to the prediction of a new subject. Main results. Extensive experiments are carried out on three EEG-based MI datasets, i.e. brain–computer interface (BCI) competition IV dataset IIb, BCI competition III dataset IVa, and BCI competition IV dataset I, and the average accuracy reaches to 86.42%, 81.71% and 79.35% on the three datasets respectively. Furthermore, the statistical analysis also indicates the significant performance improvement of SACNN. Significance. This paper reveals the importance of the temporal-spatial features on EEG-based MI classification task. Our proposed SACNN model can make fully use of the temporal-spatial information to achieve the purpose.

Privacy-Preserving Domain Adaptation for Motor Imagery-Based Brain-Computer Interfaces.

Electroencephalogram (EEG) is one of the most widely used signals in motor imagery (MI) based brain-computer interfaces (BCIs). Domain adaptation has been frequently used to improve the accuracy of EEG-based BCIs for a new user (target domain), by making use of labeled data from a previous user (source domain). However, this raises privacy concerns, as EEG contains sensitive health and mental information. It is very important to perform privacy-preserving domain adaptation, which simultaneously improves the classification accuracy for a new user and protects the privacy of a previous user. We propose augmentation-based source-free adaptation (ASFA), which consists of two parts: 1) source model training, where a novel data augmentation approach is proposed for MI EEG signals to improve the cross-subject generalization performance of the source model; and, 2) target model training, which simultaneously considers uncertainty reduction for domain adaptation and consistency regularization for robustness. ASFA only needs access to the source model parameters, instead of the raw EEG data, thus protecting the privacy of the source domain. We further extend ASFA to a stricter privacy-preserving scenario, where the source model's parameters are also inaccessible. Experimental results on four MI datasets demonstrated that ASFA outperformed 15 classical and state-of-the-art MI classification approaches. This is the first work on completely source-free domain adaptation for EEG-based BCIs. Our proposed ASFA achieves high classification accuracy and strong privacy protection simultaneously, important for the commercial applications of EEG-based BCIs.

Supervised and Semisupervised Manifold Embedded Knowledge Transfer in Motor Imagery-Based BCI.

A long calibration procedure limits the use in practice for a motor imagery (MI)-based brain-computer interface (BCI) system. To tackle this problem, we consider supervised and semisupervised transfer learning. However, it is a challenge for them to cope with high intersession/subject variability in the MI electroencephalographic (EEG) signals. Based on the framework of unsupervised manifold embedded knowledge transfer (MEKT), we propose a supervised MEKT algorithm (sMEKT) and a semisupervised MEKT algorithm (ssMEKT), respectively. sMEKT only has limited labelled samples from a target subject and abundant labelled samples from multiple source subjects. Compared to sMEKT, ssMEKT adds comparably more unlabelled samples from the target subject. After performing Riemannian alignment (RA) and tangent space mapping (TSM), both sMEKT and ssMEKT execute domain adaptation to shorten the differences among subjects. During domain adaptation, to make use of the available samples, two algorithms preserve the source domain discriminability, and ssMEKT preserves the geometric structure embedded in the labelled and unlabelled target domains. Moreover, to obtain a subject-specific classifier, sMEKT minimizes the joint probability distribution shift between the labelled target and source domains, whereas ssMEKT performs the joint probability distribution shift minimization between the unlabelled target domain and all labelled domains. Experimental results on two publicly available MI datasets demonstrate that our algorithms outperform the six competing algorithms, where the sizes of labelled and unlabelled target domains are variable. Especially for the target subjects with 10 labelled samples and 270/190 unlabelled samples, ssMEKT shows 5.27% and 2.69% increase in average accuracy on the two abovementioned datasets compared to the previous best semisupervised transfer learning algorithm (RA-regularized common spatial patterns-weighted adaptation regularization, RA-RCSP-wAR), respectively. Therefore, our algorithms can effectively reduce the need of labelled samples for the target subject, which is of importance for the MI-based BCI application.

Novel channel selection model based on graph convolutional network for motor imagery.

Multi-channel electroencephalography (EEG) is used to capture features associated with motor imagery (MI) based brain-computer interface (BCI) with a wide spatial coverage across the scalp. However, redundant EEG channels are not conducive to improving BCI performance. Therefore, removing irrelevant channels can help improve the classification performance of BCI systems. We present a new method for identifying relevant EEG channels. Our method is based on the assumption that useful channels share related information and that this can be measured by inter-channel connectivity. Specifically, we treat all candidate EEG channels as a graph and define channel selection as the problem of node classification on a graph. Then we design a graph convolutional neural network (GCN) model for channels classification. Channels are selected based on the outputs of our GCN model. We evaluate our proposed GCN-based channel selection (GCN-CS) method on three MI datasets. On three datasets, GCN-CS achieves performance improvements by reducing the number of channels. Specifically, we achieve classification accuracies of 79.76% on Dataset 1, 89.14% on Dataset 2 and 87.96% on Dataset 3, which outperform competing methods significantly.

A Parallel Feature Fusion Network Combining GRU and CNN for Motor Imagery EEG Decoding.

In recent years, deep-learning-based motor imagery (MI) electroencephalography (EEG) decoding methods have shown great potential in the field of the brain–computer interface (BCI). The existing literature is relatively mature in decoding methods for two classes of MI tasks. However, with the increase in MI task classes, decoding studies for four classes of MI tasks need to be further explored. In addition, it is difficult to obtain large-scale EEG datasets. When the training data are limited, deep-learning-based decoding models are prone to problems such as overfitting and poor robustness. In this study, we design a data augmentation method for MI-EEG. The original EEG is slid along the time axis and reconstructed to expand the size of the dataset. Second, we combine the gated recurrent unit (GRU) and convolutional neural network (CNN) to construct a parallel-structured feature fusion network to decode four classes of MI tasks. The parallel structure can avoid temporal, frequency and spatial features interfering with each other. Experimenting on the well-known four-class MI dataset BCI Competition IV 2a shows a global average classification accuracy of 80.7% and a kappa value of 0.74. The proposed method improves the robustness of deep learning to decode small-scale EEG datasets and alleviates the overfitting phenomenon caused by insufficient data. The method can be applied to BCI systems with a small amount of daily recorded data.

SPD-CNN: A plain CNN-based model using the symmetric positive definite matrices for cross-subject EEG classification with meta-transfer-learning

The electroencephalography (EEG) signals are easily contaminated by various artifacts and noise, which induces a domain shift in each subject and significant pattern variability among different subjects. Therefore, it hinders the improvement of EEG classification accuracy in the cross-subject learning scenario. Convolutional neural networks (CNNs) have been extensively applied to EEG-based Brain-Computer Interfaces (BCIs) by virtue of the capability of performing automatic feature extraction and classification. However, they have been mainly applied to the within-subject classification which would consume lots of time for training and calibration. Thus, it limits the further applications of CNNs in BCIs. In order to build a robust classification algorithm for a calibration-less BCI system, we propose an end-to-end model that transforms the EEG signals into symmetric positive definite (SPD) matrices and captures the features of SPD matrices by using a CNN. To avoid the time-consuming calibration and ensure the application of the proposed model, we use the meta-transfer-learning (MTL) method to learn the essential features from different subjects. We validate our model by making extensive experiments on three public motor-imagery datasets. The experimental results demonstrate the effectiveness of our proposed method in the cross-subject learning scenario.