Application of Convolutional Neural Network Method in Brain Computer Interface

Abstract

Pattern Recognition is the most important part of the brain computer interface (BCI) system. More and more profound learning methods were applied in BCI to increase the overall quality of pattern recognition accuracy, especially in the BCI based on Electroencephalogram (EEG) signal. Convolutional Neural Networks (CNN) holds great promises, which has been extensively employed for feature classification in BCI. This paper will review the application of the CNN method in BCI based on various EEG signals.

Full text

Journal of Physics: Conference Series
PAPER • OPEN ACCESS
Application of Convolutional Neural Network
Method in Brain Computer Interface
To cite this article: Lingzhi Chen et al 2021 J. Phys.: Conf. Ser. 2078 012044
View the article online for updates and enhancements.
You may also like
Spatial co-adaptation of cortical control
columns in a micro-ECoG brain–computer
interface
A G Rouse, J J Williams, J J Wheeler et al.
-
Expanding the (kaleido)scope: exploring
current literature trends for translating
electroencephalography (EEG) based
brain–computer interfaces for motor
rehabilitation in children
E Kinney-Lang, B Auyeung and J
Escudero
-
An artificial intelligence that increases
simulated brain–computer interface
performance
Sebastian Olsen, Jianwei Zhang, Ken-Fu
Liang et al.
-
This content was downloaded from IP address 181.215.87.142 on 17/11/2021 at 06:02

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd
ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
1
Application of Convolutional Neural Network Method in
Brain Computer Interface
Lingzhi Chen1, a, *, †, Wei Deng2, b, *,† and Chunjin Ji3, c, *,†
1 Computer Science, University of Nottingham, Nottingham, 999020, England
2 Biomedical Engineering, Tianjin Medical University, Tianjin, 300000, China
3 Mechanical design and manufacture and automation, Tianjin Technology and
Education University, Tianjin, 300000, China
*Corresponding author’s e-mail: a scylc2@nottingham.edu.cn,b576552359@qq.com,
c jcj19991010@163.com
†These authors contributed equally.
Abstract. Pattern Recognition is the most important part of the brain computer interface
(BCI) system. More and more profound learning methods were applied in BCI to
increase the overall quality of pattern recognition accuracy, especially in the BCI
based on Electroencephalogram (EEG) signal. Convolutional Neural Networks (CNN)
holds great promises, which has been extensively employed for feature classification
in BCI. This paper will review the application of the CNN method in BCI based on
various EEG signals.
1. Introduction
In decades, brain computer interface (BCI) has already become a investigation hotspot in the
following fields, including brain science, rehabilitation engineering, biomedical engineering, and
automatic control of human-machine. The reason is that BCI does not need to rely on the peripheral
nerves and muscles of the brain to output information but establishes direct communication channels
to make up for some defects or deficiencies of human beings. BCI system involves signal acquisition,
feature classification and control devices, and so on, among which pattern recognition may stand out.
Deep learning is the method to realize the internal procedures and interpretation levels of sample data
and as-obtained information during learning, which is conductive to interpret data. The outcome is
that the tested machine can get the goal which is similar to being analytical and created by human
beings. Not only can it handle texts, but it also recognizes the images or sounds. More and more deep
learning methods were applied in BCI to improve the overall quality of pattern recognition accuracy,
especially in the BCI based on Electroencephalogram (EEG) signal.
Convolutional Neural Networks (CNN) is a critical class of feedforward neural network among all
those deep learning models. It includes convolutional calculation and has deep structure, widely
applied to BCI for feature extraction and classification in BCI.
This paper will review the traditional BCI feature classification method and other machine
learning methods, especially neural network models like CNN and RNN. We will mostly focus on the
application of these classification methods in BCI pattern recognition.

ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
2
2. BCI system
Note that Brain-Computer Interface (BCI) is able to interpret activity forms of the human brain into
communications or instructions to interconnect with the outer world [1]. The major limitation of BCI
is to accurately identify human intentions at weak signal-to-noise ratios (SNR) of brain signals.
2.1 Pattern Recognition
Despite the huge achievements of conventional BCI systems, it is still tough to develop BCI. Various
biological and environmental artifacts easily corrupt brain signals. Whereas some signal
preprocessing and feature selection and extraction techniques have been established, feature
engineering is extremely dependent on human expertise in a particular field [2].
2.2 Interface Application
The applications of the BCI system may replace the natural output that is lost due to injury or disease, such
as a person who has lost the ability to speak by writing through a BCI or by speaking through a speech
synthesizer. Secondly, the output of the brain-computer interface can restore lost function. For example,
cochlear implants have helped hundreds of thousands of deaf patients reestablish hearing, and synthetic
eyeballs is able to assist blind patients to see again, etc. Brain computer interface (BCI) technology can
also train the brain's motor cortex to help stroke patients recover after they have lost control of their limbs.
Furthermore, mainly for healthy people, to achieve functional expansion. In engineering psychology,
cognitive load, fatigue degree and other states of personnel in special work positions. They play an
incredibly crucial role in job performance and job safety. In education, brain-computer interfaces do matter
in assessing students' dynamic attention level and optimize teachers' education arrangement.
3. BCI feature classification method
This section presents a thorough introduction of pattern recognition methods used in BCI systems.
Figure 1 demonstrates a taxonomy of the BCI feature classification method, including EEG feature
extraction, machine learning method, and deep learning method, namely Neural Network Model.
EEG feature extraction (Section 3.1); Machine learning method includes Decision Trees, Naive
Bayesian Classification and Support Vector Machine (Section 3.2); Neural network model including
Convolutional Neural Networks, Recurrent Neural Network and Bi-directional LSTM RNN (Section
3.3).

ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
3
Fig.1. The BCI feature classification method is used in the BCI system.
3.1 EEG feature extraction
In the field of BCI, it is highly desired to examine data with many differences. If each indicator is
investigated separately, the assessment tends to be unique. Therefore, the following common feature
extraction methods are generated.
3.1.1 Principal Component Analysis (PCA)
PCA is the most broadly applied data dimension reduction algorithm. Celine uses an online database
of active EEG signals containing left and right-hand motion images (Glass dataset B). These data are
studied with extracted features by PCA.[4] In this case, MI information. The dimensionality reduction
of the signal can be achieved by using the projection method to create a suite of linear/nonlinear
conversions of the input variables. Anticipated MI outcomes were one hundred percent faithful to the
supposed classification, i.e., right-handed MI executions were classified as Category 2, and left-
handed MI executions were classified as Category 1. Through the experiment, the accuracy is 100%.
The obtained signal has sufficient correlation to ensure that the motor image is correctly classified
using the LR method, but the size is reduced.
3.1.2 Independent Component Analysis (ICA)
ICA is a calculation means to separate multiple signals into additive sub components. This is
accomplished by assuming that the subcomponents are non-Gaussian and statistically self-
determining of each other. PCA shows that the principal components are orthogonal to each other,
and the samples are Gaussian distribution; ICA does not require samples to be Gaussian distribution.
Nguyen utilized neurophysiological to estimate whether auditory cue were able to acquire identical
data compared to visual did and denoise signal through band pass filter.[5]
BCI Featrue
Classification
Method
EEG Feature
Extraction
Principal
Component Analysis
(PCA)
Independent
Component Analysis
(ICA)
Canonical
Correlation Analysis
(CCA)
Machine Learning
Method
Decision Trees
Naive Bayesian
Classification
Support Vector
Machine (SVM)
Neural Network
Model
Convolutional
Neural Networks
(CNN)
Recurrent Neural
Network (RNN)
Bi-directional LSTM
RNN (BRNN)

ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
4
3.1.3 Canonical Correlation Analysis (CCA)
CCA is to correlate a linear relationship between 2 multidimensional variables. CCA serves as
complicated tags to direct feature selection to the underlying semantics. Shao successfully applied
CCA and its extension method to frequency identification of BCI system based on SSEP and
proposed TCCA and an improved filter bank frequency identification method based on TCCA, Filter
Bank Time Local Specification Correlation Analysis (FBTCCA).[6] Lin et al. used the SSEP data set
of 10 healthy subjects (three women, aged eighteen to thirty-one, all right-handed) to participate in the
offline test. The time information of the CCA method is considered for the first time to reduce the
artifacts and improve the algorithm's accuracy under a short time window. Combined with the filter
library and TCCA to improve the classification accuracy, the final accuracy is 91.16% when the
processing time is 1.5s.
3.2 Machine learning method for feature classification
Machine learning requires a computer and contributes to model real-time human learning. Since the
1950s, artificial intelligence (AI) has experienced the "reasoning period", "knowledge period",
"knowledge engineering bottleneck," and "machine learning period". This has led to various machine
learning methods, three of which are common in the following sections.
3.2.1 Decision Trees
Decision Trees are in the light of the identified probability of manifestation of different circumstances.
Joadder proposed a new method of computer-assisted feature selection to clarify the most excellent
feature set for differentiating motor images instead of the manual feature selection used in previous
studies. [7] A decision tree classifies the functions selected by this method to verify the overall
performance. During the process of recording, the subjects were requested to conduct a motion image
task of their left, right, or right foot, but the competition only provided clues for the right and right
foot categories [8]. There are 280 trials per subject, of which there are 140 trials per subject. Using a
single feature is the most accurate feature "combination" in the test object. However, despite these
results, the algorithms used still have independent success due to their precision and the reduced
computational load.
3.2.2 Naive Bayesian classification
The Bayesian technique is in line with Bayesian theory and categorizes the dataset by probability and
statistics. Then, Naive Bayes approach is a simplified method according to Bayes algorithm.
3.2.3 Support Vector Machine (SVM)
SVM represents one regulated learning model associated with concerned learning algorithms in
machine learning. One SVM training algorithm determines a model and allocates new instances to
one class or other classes to make it a nonprobabilistic binary linear classification. Its core is the
kernel function. After establishing the kernel function, since the recognized data of determining the
kernel function also have a variety of inaccuracies, considering the problem of generalization, two-
parameter variables.
Ghumman proposed a feature classification method in line with SVM and improved its
performance by optimizing polynomial kernel parameters.[9] In accordance with the standardized
global EEG 10-20 system, electrodes are positioned on different parts of the scalp to record the brain's
electrical activity.[10] These signals from the electrodes reflect the subject's motion image (MI)
activity.[11] Then, EEG signals were recorded while performing various MI tasks such as hand, foot,
and tongue movements.[12] A polynomial kernel (SVM-PK) support vector machine method for EEG
signal classification in a BCI system on the basis of MI is proposed in this work. Using the network
search means to select the best value of the polynomial kernel, the performance is improved. This
experiment is to select the kernel and then optimize it. At the same time, the steps of regularization

ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
5
parameter (C) were changed from 0.1 to 100[0.1,10,20,-,90,100] by coarse grid search. Under all
these C values, the classification accuracy reaches 0.664.
4. Neural network Model
Deep learning is one rising technique that utilizes neural networks to transform or represent the input
nonlinearly. Deep neural networks (DNN) are the foundation of deep learning. To understand DNN,
we must first understand the DNN model. Through the construction of deep neural networks,
numerous analysis activities are carried out.
4.1 Convolutional Neural Networks
Under the framework of the biology of the visual cortex, the idea that specific elements of the system
have specific tasks is applied to the machine, which is the basis of CNN. The reason why CNN can
work is that computer can categorize images by looking for low-level features, and then build more
abstract concepts through a series of convolutional levels.
Fig.2. The basic structure of CNN.
Nijisha Shajil et al. used CNN for the classification of MI signals.[13] EEG signals are obtained
through 16 channels and filtered using a bandpass filter with a frequency range of 1 to 100 Hz. The
spectrogram of the spatially filtered signal was provided to CNN as input. Monoconvolutional layer
CNN is constructed for classifying MI EEG signals of the left hand, right hand, hands, and feet.
4.2 Recurrent Neural Network
In the CNN mentioned above, the input and output of training samples are relatively determined. The
idea of RNN is to use sequence information. It is proposed that all inputs and outputs are independent
of each other. But for many other tasks, this idea is very bad. If you need to predict the next word in a
sentence, it's helpful to know the preceding word. In theory, RNN can use the information of any
length sequence, but it can only deal with very limited information of the first few steps in practice.
Sumanto Dutta has proposed a new data-enhancement approach to address the challenges posed by
the scarcity of EEG data for training deep learning models (e.g., RNN).[14] Flow EEG data were
collected from 16 participants using the Emotiv EPOC+ 14-channel headset. To sort out EEG data,
the system using NVIDIA's GeForce Titan XP GPU has 12GB of memory to implement the deep
learning algorithm. The verification accuracy of the enhanced data is better than that of the
unenhanced data. Solid experimental results indicates that the performance of the mental state
estimator is improved by data enhancement, and the accuracy rate is 98%.
4.3 Bi-directional LSTM RNN
Normally, RNN is only able to foresee the output of the next time in proportion to the time sequence
information of the preceding time. Still, the current time's output is related to the original state and
may be associated with the future state. BRNN has two RNN superimposed together, and the states of
the two RNN determine the output. The basic idea of BRNN is to propose that every training
sequence is two RNN forward and backward, and these two RNN are connected with an output layer.
Cai, CNN, and BRNN were combined for facial expression recognition.[15] Cheavd, the challenge
data set, consists of 140 minutes of spontaneous emotion clips extracted from movies, TV shows, and

ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
6
talk shows. The database includes 238 speakers, ranging from children to the elderly. It consists of
eight emotions: anger, anxiety, disgust, happiness, neutrality, sadness, surprise, and worry. In contrast,
the dataset is extended using two forms of data enhancement. This method acts well in recognizing
emotions such as anger (74.07%), happiness (46.27%), surprise (54.48%), worry (33.33%), anger
(22.37%), and sadness (26.88%), suggesting its big role in multi-emotion recognition. The highest
recognition rate was 74.07%, compared to only 4.76% at baseline, demonstrating that this method
seems to be very effective in identifying aversive manifestations. However, the neutral recognition
rate was 27.51%, lower than baseline (61.97%), nothing that the lowest recognition rate in this work
was 15.79%.
5. Convolutional Neural Networks in BCI
The commonly used BCI-EEG paradigm includes visual evoked potential, motor imagination,
emotion, and so on. Next, the investigations in every BCI signal and the deep learning method related
to the feature classification will be summarized.
5.1 Application in Visual Evoked Potentials based BCI
5.1.1VEP
In general, the ERP signals are examined by the P300 phenomenon. Cecotti et al.[16]attempted to
improve the recognition correctness of P300 to obtain more accurate word spellings. A new-found
model in line with CNN is proposed. The model includes five low-level CNN classifiers with various
feature sets, and the low-level classifiers vote the final high-level results. In the third BCI competition,
the highest accuracy of Dataset II was 95.5%. Liu and co-workers [17] created a Batch Normalized
Neural Network (BN3), a variant of CNN in the P300 speller. The method is divided into six layers,
and each batch is normalized. Maddula et al. [18] filtered the P300 signals with visual stimulation
through a bandpass (2~35Hz).
5.1.2 SSVEP
Most deep learning-based researches in SSEP concentrate on SSVEP like [19, 20]. Waytowich et al.
[21] applied a compact CNN model to process raw SSVEP signals directly without any hand-crafted
features. The average accuracy of the reported cross-subjects was about 80%. Atia et al.[22] aimed to
determine a suitable intermediate interpretation of SSVEP. Aznan et al.[23] explored SSVEP
classification of signals collected through dry electrodes. Thomas et al.[24] The original SSVEP
signal is triggered by a band signal (5~48Hz), and then the discrete FFT is operated on 512
consecutive points. The processed data were independently classified by CNN (69.03%) and LSTM
(66.89%). Gao et al. [25] designed a SSMVEP signal-based trolley control system and introduced a
deep learning method. They then constructed a convolutional neural network (CNN-LSTM)
framework with a long and short memory.
5.2 Application in Motor Imagine based NCI
Extreme Learning Machine (ELM )[26] has shown superiorities on the classification of MI EEG and
real-motor EEG [27, 28]. Uktveris et al.[29]extracted many EEG features, such as mean channel
energy (MCE), mean window energy (MWE), channel variance (MWE, CV), mean band power, etc.
All extracted features are sent to 2D CNN for classification. Lee et al.[30] deal with the MI EEG
signals by wavelet transform and manually extracted the PSD from the Mu and Beta bands for the
first time. Wang et al.[31] designed a fast convolutional feature extraction method based on CNN to
learn potential features from MI-EEG signals. Several weak classifiers are applied to select important
features for the final classification. Hartman et al.[27] studied EEG signals induced by real motor
effects. They studied how CNN represented spectral features through network intermediate phase
sequences, showing high sensitivity to EEG phase features at the early stage and high sensitivity to
EEG amplitude features later. Some studies have proposed a mixed model of electroencephalography

ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
7
for the identification of myocardial infarction[32].Tan et al.[33] proposed a complex system to
achieve multimodal EEG classification. DE-noising AE was used to reduce dimension. A method
combining multiple views CNN and RNN is proposed to discover the potential temporal and
spatiotemporal information of low dimensional representations. They obtained the IIA dataset from
BCI Competition IV with an average accuracy of 72.22%.Fadel et al.[34]proposed a chessboard
image transformation method, which converts EEG signals of moving images into images for
classification using hybrid deep learning models. The network model is composed of DCNN to
extract spatial and frequency features, LSTM to extract time features, and finally, divided into five
categories (four motion image tasks and one motion image task).The results are encouraging, and the
classification accuracy of checkerboard method is 68.72%. Lun et al.[35] proposed a deep CNN
structure with time and space filter separation. Echtioui et al.[36]compared two models, CNN and
LSTM, on the same basic data set, an optimal classification method based on the deep learning
method is proposed. The BCI Competition IV dataset 2A was used as the base dataset to test both
classification methods.
5.3 Application in Emotional BCI
Of great note, one person’s emotion could be identified by three factors, namely valence, arousal, and
dominance, where each one can be rated by an integer between 1 to 9 or recognized positive and
negative [37, 38]. Wang et al. [39] used the CNN algorithm to classify single EEGs. Interestingly,
they enhanced the training set by adding Gaussian noise to the original sample to generate new EEG
samples. Li et al.[40] proposed a new hierarchical convolutional neural network (HCNN) to identify
subjects' emotions (positive, neutral, negative) and achieved an accuracy rate of 88.2%.In the HCNN
structure, each convolution kernel has only a local acceptance field. The kernel can capture the
correlation between adjacent electrodes, which may be of great value for the recognition task.
Sheykhivand et al.[41]proposed an automatic two-stage (negative, positive) and three-stage (negative,
positive, neutral) classification of emotions based on EEG signals. In this method, EEG raw signals
are directly applied to Convolutional Neural Network (CNN-LSTM) and Short and Long Term
Memory Network (CNN-LSTM).
6 Conclusion
In this paper, we introduce some data processing methods in machine learning and focus on the latest
application of CNN in deep learning. Compared with traditional methods, CNN can improve the accuracy
of image feature extraction. Through the flexible use of the convolutional layer and the combination of the
convolutional layer, the basic structure of CNN can be transformed to improve the processing accuracy.
The combination of CNN and LSTM has a prominent effect of improving the processing accuracy in MI,
SSMVEP, and Emotional directions. In this paper, we introduce some data processing methods in machine
learning and focus on the latest application of CNN in deep learning. Compared with traditional methods,
CNN can improve the accuracy of image feature extraction. Through the flexible use of the convolutional
layer and the combination of convolutional layer, the basic structure of CNN can be transformed to
improve. The combination of CNN and LSTM has a prominent effect of improving the processing
accuracy in MI, SSMVEP, and Emotional directions.
References
[1]Fabien Lotte, Laurent Bougrain, Andrzej Cichocki, Maureen Clerc, Marco Congedo, Alain
Rakotomamonjy, and Florian Yger. 2018. A review of classification algorithms for EEG-
based brain–computer interfaces: a 10 year update. Journal of neural engineering 15, 3 (2018)
[2]Xiang Zhang, Lina Yao, Xianzhi Wang, Jessica Monaghan, and David McAlpine. 2018. A Survey
on Deep LearningbasedBrainComputerInterface:RecentAdvancesandNewFrontiers.
1,1,Article1(January2018), 66 pages.
[3]HubertCecotti,MiguelPEckstein,andBarryGiesbrecht.2014.Single-trialclassificationofevent-

ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
8
relatedpotentials in rapid serial visual presentation tasks using supervised spatial filtering.
IEEE transactions on neural networks and learning systems 25, 11 (2014), 2030–2042.
[4]Soeiro C.F.C. (2019) EEG Signal Analysis Using PCA and Logistic Regression. In: Costa-Felix R.,
Machado J., Alvarenga A. (eds) XXVI Brazilian Congress on Biomedical Engineering.
IFMBE Proceedings, vol 70/2. Springer, Singapore.
[5]Nguyen T.H., Park SM., Ko KE., Sim KB. (2012) Improvement of Spatial Filtering by Using ICA
in Auditory Stimuli BCI Systems of Hand Movement. In: Lee G., Howard D., Kang J.J.,
Ślęzak D. (eds) Convergence and Hybrid Information Technology. ICHIT 2012. Lecture
Notes in Computer Science, vol 7425. Springer, Berlin, Heidelberg.
[6]Shao, X., Lin, M. Filter bank temporally local canonical correlation analysis for short time window
SSVEPs classification. Cogn Neurodyn 14, 689–696 (2020).
[7]Joadder, M.A.M., Myszewski, J.J., Rahman, M.H. et al. A performance based feature selection
technique for subject independent MI based BCI. Health Inf Sci Syst 7, 15 (2019).
[8]Dornhege G, Blankertz B, Curio G, Muller KR. To improve the bit rate of single trial classification
in non-invasive EEG by feature combination and multi-class paradigm. IEEE Trans Biomed
Eng. 2004; 51 (6) : 993-1002.
[9]Ghumman, M.K., Singh, S., Singh, N. et al. Optimization of parameters for improving the
performance of EEG-based BCI system. J Reliable Intell Environ 7, 145–156 (2021).
[10]Costantini G, Todisco M, Cassari D, Carlota M, Saggio G, Bianchi L, Abafati M, Kitadamo L
(2009) SVM classification of electronic signals in brain-computer interfaces. In: Proceedings
of the 2009 Conference on Neural Networks WIRN09: Proceedings of the 19th Italian
Symposium on Neural Networks, Nanvieri, Salerno, Italy, 28 May 2009 solstice 30. IOS
Press, pp. 229-233.
[11]Lotte F, Bougrain L, Cichocki A, Clerc M, Conedo M, Rakotomamonjy A, Yger F (2018) A
review of EEG-based classification algorithms for brain-computer interfaces: A 10-year
update. J Neural Eng 15(3):031005.
[12]Gaur P, Pachori RB, Wang H, Prasad G (2018) An EEG-based multi-class BCI classification
using filtering and Riemannian geometry based on multivariate empirical mode
decomposition. Application of expert systems 95:201-211.
[13]Nijisha Shajil, Sasikala Mohan, Poonguzhali Srinivasan, Janani Arivudaiyanambi &Arunnagiri
Arasappan Murrugesan, A BCI-based application uses convolutional neural network to
classify EEG signals from spatial filter motor images, Journal of Medical and Biological
Engineering volume 40, pages663–672(2020).
[14]Dutta S., Nandy A. (2019) Data Augmentation for Ambulatory EEG Based Cognitive State
Taxonomy System with RNN-LSTM. In: Bramer M., Petridis M. (eds) Artificial Intelligence
XXXVI. SGAI 2019. Lecture Notes in Computer Science, vol 11927. Springer, Cham.
[15]Cai Y., Zheng W., Zhang T., Li Q., Cui Z., Ye J. (2016) Video Based Emotion Recognition Using
CNN and BRNN. In: Tan T., Li X., Chen X., Zhou J., Yang J., Cheng H. (eds) Pattern
Recognition. CCPR 2016. Communications in Computer and Information Science, vol 663.
Springer, Singapore.
[16Cecotti, H.J.C.P.I.E.M.B.S., Convolutional neural networks for event-related potential detection:
impact of the architecture. 2017: p. 2031-2034.
[17]Liu, et al., Deep learning based on Batch Normalization for P300 signal detection. 2018.
[18]Maddula, R.K., Going Deeper with Recurrent Convolutional Neural Networks for
ClassifyingP300 BCI Signals. 2017.
[19]Ahn, M.H. and B.K. Min. Applying deep-learning to a top-down SSVEP BMI. in International
Conference on Brain-Computer Interface.
[20]Kwak, N.S., K. Müller, and S.W.J.P.O. Lee, A convolutional neural network for steady state
visual evoked potential classification under ambulatory environment. 2017. 12.
[21]Waytowich, N., et al., Compact Convolutional Neural Networks for Classification of
Asynchronous Steady-state Visual Evoked Potentials. 2018. 15(6): p. 066031.1-066031.13.

ICAITA 2021
Journal of Physics: Conference Series 2078 (2021) 012044
IOP Publishing
doi:10.1088/1742-6596/2078/1/012044
9
[22]Attia, M., et al. A time domain classification of steady-state visual evoked potentials using deep
recurrent-convolutional neural networks. in 2018 IEEE 15th International Symposium on
Biomedical Imaging (ISBI 2018). 2018.
[23]Aznan, N., et al., On the Classification of SSVEP-Based Dry-EEG Signals via Convolutional
Neural Networks.
[24]Thomas, J., et al. Deep learning-based classification for brain-computer interfaces. in IEEE
International Conference on Systems. 2017.
[25]A Deep Learning Method for Improving the Classification Accuracy of SSMVEP-based BCI %J
Circuits and Systems II: Express Briefs, IEEE Transactions on. 2020. PP(99): p. 1-1.
[26]Duan, L., et al., Motor Imagery EEG Classification Based on Kernel Hierarchical Extreme
Learning Machine. 2017.
[27]Hartmann, K.G., R.T. Schirrmeister, and T. Ball, Hierarchical internal representation of spectral
features in deep convolutional networks trained for EEG decoding. 2017.
[28]Nurse, E., et al., Decoding EEG and LFP signals using deep learning: heading TrueNorth. 2016.
p. 259-266.
[29]Uktveris, T., V.J.I.t. Jusas, and control, Application of Convolutional Neural Networks to Four-
Class Motor Imagery Classification Problem. 2017. 46(2).
[30]Lee, H.K. and Y.S. Choi. A convolution neural networks scheme for classification of motor
imagery EEG based on wavelet time-frequecy image. in 2018 International Conference on
Information Networking (ICOIN). 2018.
[31]Qian, W., Y. Hu, and C.J.I.S.o.N.N. He, Multi-channel EEG Classification Based on Fast
Convolutional Feature Extraction. 2017.
[32]Dai, M., et al., EEG Classification of Motor Imagery Using a Novel Deep Learning Framework.
2019. 19(3).
[33]Tan, C., et al. Multimodal Classification with Deep Convolutional-Recurrent Neural Networks for
Electroencephalography. in International Conference on Neural Information Processing.
2017.
[34]Fadel, W., et al. Chessboard EEG Images Classification for BCI Systems Using Deep Neural
Network. in International Conference on Bio-inspired Information and Communication
Technologies. 2020. Springer.
[35]Lun, X., et al., A Simplified CNN Classification Method for MI-EEG via the Electrode Pairs
Signals. 2020.
[36]Echtioui, A., et al., Multi-class Motor Imagery EEG Classification using Convolution Neural
Network. 2021. p. 591-595.
[37]Li, J., Z. Zhang, and H. He. Implementation of EEG Emotion Recognition System Based on
Hierarchical Convolutional Neural Networks. in International Conference on Brain Inspired
Cognitive Systems. 2016.
[38]Jiang, H., W. Liu, and L. Yao. Analyze EEG Signals with Convolutional Neural Network Based on
Power Spectrum Feature Selection. in International Conference on Computer Engineering &
Networks. 2017.
[39]Fang, W., et al. Data Augmentation for EEG-Based Emotion Recognition with Deep
Convolutional Neural Networks. in International Conference on Multimedia Modeling. 2018.
[40]Li, J., Z. Zhang, and H.J.C.C. He, Hierarchical Convolutional Neural Networks for EEG-Based
Emotion Recognition. 2017.
[41]Sheykhivand, S., et al., Recognizing Emotions Evoked by Music using CNN-LSTM Networks on
EEG signals. 2020. PP(99): p. 1-1.