EEGNet-based multi-source domain filter for BCI transfer learning

Deep learning has great potential on decoding EEG in brain-computer interface. While common deep learning algorithms cannot directly train models with data from multiple individuals because of the inter-individual differences in EEG. Collecting enough data for each subject to satisfy the training of deep learning would result in an increase in training cost. This study proposes a novel transfer learning, EEGNet-based multi-source domain filter for transfer learning (EEGNet-MDFTL), to reduce the amount of training data and improve the performance of BCI. The EEGNet-MDFTL uses bagging ensemble learning to learn domain-invariant features from the multi-source domain and utilizes model loss value to filter the multi-source domain. Compared with baseline methods, the accuracy of the EEGNet-MDFTL reaches 91.96%, higher than two state-of-the-art methods, which demonstrates source domain filter can select similar source domains to improve the accuracy of the model, and remains a high level even when the data amount is reduced to 1/8, proving that ensemble learning learns enough domain invariant features from the multi-source domain to make the model insensitive to data amount. The proposed EEGNet-MDFTL is effective in improving the decoding performance with a small amount of data, which is helpful to save the BCI training cost.

Cross-subject spatial filter transfer method for SSVEP-EEG feature recognition

Objective.Steady-state visual evoked potential (SSVEP) is an important control method of the brain-computer interface (BCI) system. The development of an efficient SSVEP feature decoding algorithm is the core issue in SSVEP-BCI. It has been proposed to use user training data to reduce the spontaneous electroencephalogram activity interference on SSVEP response, thereby improving the feature recognition accuracy of the SSVEP signal. Nevertheless, the tedious data collection process increases the mental fatigue of the user and severely affects the applicability of the BCI system.Approach.A cross-subject spatial filter transfer (CSSFT) method that transfer the existing user model with good SSVEP response to the new user test data without collecting any training data from the new user is proposed.Main results.Experimental results demonstrate that the transfer model increases the distinction of the feature discriminant coefficient between the gaze following target and the non-gaze following target and accurately identifies the wrong target in the fundamental algorithm model. The public datasets show that the CSSFT method significantly increases the recognition performance of canonical correlation analysis (CCA) and filter bank CCA. Additionally, when the data used to calculate the transfer model contains one data block only, the CSSFT method retains its effective feature recognition capabilities.Significance.The proposed method requires no tedious data calibration process for new users, provides an effective technical solution for the transfer of the cross-subject model, and has potential application value for promoting the application of the BCI system.

Neural Decoding of Chinese Sign Language With Machine Learning for Brain-Computer Interfaces

Limb motion decoding is an important part of brain-computer interface (BCI) research. Among the limb motion, sign language not only contains rich semantic information and abundant maneuverable actions but also provides different executable commands. However, many researchers focus on decoding the gross motor skills, such as the decoding of ordinary motor imagery or simple upper limb movements. Here we explored the neural features and decoding of Chinese sign language from electroencephalograph (EEG) signal with motor imagery and motor execution. Sign language not only contains rich semantic information, but also has abundant maneuverable actions, and provides us with more different executable commands. In this paper, twenty subjects were instructed to perform movement execution and movement imagery based on Chinese sign language. Seven classifiers are employed to classify the selected features of sign language EEG. L1 regularization is used to learn and select features that contain more information from the mean, power spectral density, sample entropy, and brain network connectivity. The best average classification accuracy of the classifier is 89.90% (imagery sign language is 83.40%). These results have shown the feasibility of decoding between different sign languages. The source location reveals that the neural circuits involved in sign language are related to the visual contact area and the pre-movement area. Experimental evaluation shows that the proposed decoding strategy based on sign language can obtain outstanding classification results, which provides a certain reference value for the subsequent research of limb decoding based on sign language.

Hysteresis Modeling in Iron-Dominated Magnets Based on a Multi-Layered Narx Neural Network Approach

A full-fledged neural network modeling, based on a Multi-layered Nonlinear Autoregressive Exogenous Neural Network (NARX) architecture, is proposed for quasi-static and dynamic hysteresis loops, one of the most challenging topics for computational magnetism. This modeling approach overcomes drawbacks in attaining better than percent-level accuracy of classical and recent approaches for accelerator magnets, that combine hybridization of standard hysteretic models and neural network architectures. By means of an incremental procedure, different Deep Neural Network Architectures are selected, fine-tuned and tested in order to predict magnetic hysteresis in the context of electromagnets. Tests and results show that the proposed NARX architecture best fits the measured magnetic field behavior of a reference quadrupole at CERN. In particular, the proposed modeling framework leads to a percent error below 0.02% for the magnetic field prediction, thus outperforming state of the art approaches and paving a very promising way for future real time applications.

Enhancing detection of steady-state visual evoked potentials using channel ensemble method

OBJECTIVE

This study proposed and evaluated a channel ensemble approach to enhance detection of steady-state visual evoked potentials (SSVEPs).

APPROACH

Collected multi-channel electroencephalogram (EEG) signals were classified into multiple groups of new analysis signals based on correlation analysis, and each group of analysis signals contained signals from a different number of electrode channels. These groups of analysis signals were used as the input of a training-free feature extraction model, and the obtained feature coefficients were converted into feature probability values using the softmax function. The ensemble value of multiple sets of feature probability values was determined and used as the final discrimination coefficient.

MAIN RESULTS

Compared with canonical correlation analysis (CCA), likelihood ratio test (LRT), and multivariate synchronization index (MSI) analysis methods using a standard approach, the recognition accuracies of the methods using a channel ensemble approach were improved by 5.05%, 3.87%, and 3.42%, and the information transfer rates (ITRs) were improved by 6.00%, 4.61%, and 3.71%, respectively. The channel ensemble method also obtained better recognition results than the standard algorithm on the public dataset. This study validated the efficiency of the proposed method to enhance the detection of SSVEPs, demonstrating its potential use in practical brain-computer interface (BCI) systems.

SIGNIFICANCE

A SSVEP-based BCI system using a channel ensemble method could achieve high ITR, indicating great potential of this design for various applications with improved control and interaction.

Acrophobia Quantified by EEG Based on CNN Incorporating Granger Causality

The aim of this study is to quantify acrophobia and provide safety advices for high-altitude workers. Considering that acrophobia is a fuzzy quantity that cannot be accurately evaluated by conventional detection methods, we propose a comprehensive solution to quantify acrophobia. Specifically, this study simulates a virtual reality environment called High-altitude Plank Walking Challenge, which provides a safe and controlled experimental environment for subjects. Besides, a method named Granger Causality Convolutional Neural Network (GCCNN) combining convolutional neural network and Granger causality functional brain network is proposed to analyze the subjects' noninvasive scalp EEG signals. Here, the GCCNN method is used to distinguish the subjects with severe acrophobia, moderate acrophobia, and no acrophobia in a three-class classification task or no acrophobia and acrophobia in a two-class classification task. Compared with the mainstream methods, the GCCNN method achieves better classification performance, with an accuracy of 98.74% for the two-class classification task (no acrophobia versus acrophobia) and of 98.47% for the three-class classification task (no acrophobia versus moderate acrophobia versus severe acrophobia). Consequently, our proposed GCCNN method can provide more accurate quantitative results than the comparative methods, making it to be more competitive in further practical applications.

Reduction of Onset Delay in Functional Near-Infrared Spectroscopy: Prediction of HbO/HbR Signals

An intrinsic problem when using hemodynamic responses for the brain-machine interface is the slow nature of the physiological process. In this paper, a novel method that estimates the oxyhemoglobin changes caused by neuronal activations is proposed and validated. In monitoring the time responses of blood-oxygen-level-dependent signals with functional near-infrared spectroscopy (fNIRS), the early trajectories of both oxy- and deoxy-hemoglobins in their phase space are scrutinized. Furthermore, to reduce the detection time, a prediction method based upon a kernel-based recursive least squares (KRLS) algorithm is implemented. In validating the proposed approach, the fNIRS signals of finger tapping tasks measured from the left motor cortex are examined. The results show that the KRLS algorithm using the Gaussian kernel yields the best fitting for both ΔHbO (i.e., 87.5%) and ΔHbR (i.e., 85.2%) at q = 15 steps ahead (i.e., 1.63 s ahead at a sampling frequency of 9.19 Hz). This concludes that a neuronal activation can be concluded in about 0.1 s with fNIRS using prediction, which enables an almost real-time practice if combined with EEG.

Effect of the concreteness of robot motion visual stimulus on an event-related potential-based brain-computer interface

Event-related potential (ERP)-based brain-computer interface (BCI) has been widely used in robot control. Increasing the amplitude of the ERPs is key for improving the performance of ERP-based BCI. However, using images of robot motion as visual stimuli has not been studied widely. The aim of this study is to explore the concreteness of robot motion images on ERPs. Fifteen subjects used five kinds of visual spellers employing different images as visual stimuli: squares, arrows, a single kind of robot motion, multiple kinds of robot motions, and multiple kinds of robot motions with arrows. The three robot motion stimuli induced larger N200 and P300 potentials than non-robot motion stimuli. The topography shows that robot motion stimuli also evoke stronger negativities in the anterior and occipital areas. Concrete images provide more information to the subject about the robot motion, which might help the brain extract the meaning of the image more automatically. We use a support vector machine to detect the subject's intentions. There is substantial improvement in the classification performance when using robot motion images as visual stimuli, which implies that concrete visual stimuli improve the performance of the ERP-based BCI.

Monitor-Based Spiking Recurrent Network for the Representation of Complex Dynamic Patterns

Neural networks are powerful computation tools for mimicking the human brain to solve realistic problems. Since spiking neural networks are a type of brain-inspired network, called the novel spiking system, Monitor-based Spiking Recurrent network (MbSRN), is derived to learn and represent patterns in this paper. This network provides a computational framework for memorizing the targets using a simple dynamic model that maintains biological plasticity. Based on a recurrent reservoir, the MbSRN presents a mechanism called a ‘monitor’ to track the components of the state space in the training stage online and to self-sustain the complex dynamics in the testing stage. The network firing spikes are optimized to represent the target dynamics according to the accumulation of the membrane potentials of the units. Stability analysis of the monitor conducted by limiting the coefficient penalty in the loss function verifies that our network has good anti-interference performance under neuron loss and noise. The results of solving some realistic tasks show that the MbSRN not only achieves a high goodness-of-fit of the target patterns but also maintains good spiking efficiency and storage capacity.