SIGNaL PROCESSING
algorithms, architectures, arrangements, and applications
THE INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS INC.
Automatic Detection of SSVEP using
Independent Component Analysis
Marcin Koodziej, Andrzej Majkowski, Remigiusz J. Rak
Warsaw University of Technology, Institute of Theory of Electrical Engineering, Measurement and Information
Systems, Koszykowa 75 st., 00-662 Warsaw, marcin.kolodziej@ee.pw.edu.pl
Abstract A new method of automatic SSVEP detection
in the EEG signal based on Independent Components
Analysis (A-ICA) is described in the article. For the
presented method reduction of artifacts in EEG signal is not
required. Besides, the method has low computational
complexity. In order to select the best ICA decomposition,
we conducted a comparative study of known ICA
algorithms for the use in SSVEP detection. For feature
extraction from ICA components, time-frequency analysis
(STFT) was used. The algorithm was tested on real EEG
signals recorded from 5 users, for 4 stimuli frequencies. The
achieved SSVEP classification accuracy was in the range of
53-89%, for different users. The proposed algorithm can be
successfully used in SSVEP based brain-computer
interfaces.
KeywordsSteady State Visually Evoked Potentials,
SSVEP, electroencephalogparhy, EEG, Independent,
Component Analysis, ICA.
I. INTRODUCTION
Electroencephalographic potentials, especially P300,
Event-Related Desynchronization (ERD), Event-Related
Synchronization (ERS) and Steady State Visually Evoked
Potentials (SSVEP), are commonly used in brain-
computer interfaces [1]. In recent years, brain-computer
interfaces using SSVEP arouse great interest [2]-[5].
When using stimuli in the form of light flashing with
frequency f, a synchronous response of the same
frequency and harmonic frequencies 2f, 3f is expected in
EEG signal recorded from visual cortex [6],[7]. Brain-
computer interface based on SSVEP does not require
preliminary tedious user preparation and, in addition, is
characterized by a large Information Transfer Rate (ITR)
[8]. From the reasons mentioned above, it has great
practical significance. SSVEP based BCI find use in
communication, robots control, electric power steering
systems. However, SSVEP interfaces also have several
disadvantages associated with their use. First of all, they
require to focus your attention on a flickering light source.
Stimulation with flashing light can be fatiguing for a user,
and even provoke epileptic attack (synchronization of the
visual cortex). Nevertheless, efficient and effective
algorithms of processing, analysis and classification of the
EEG signal for the use in SSVEP based interfaces are
sought [9],[10].
196
SPa 2016
September 21-23rd, 2016, Pozna, POLAND
In the case of SSVEP we can a priori determine what
is sought in the EEG signal: increased power of a periodic
signal in frequencies f, 2f, 3f. For that reason, many
SSVEP detection algorithms are based on a small number
of channels and simple methods of EEG signal analysis
[11],[12]. Unfortunately, this approach does not allow us
to use the full potential of processing and analysis of the
EEG signal coming from all electrodes. EEG signals are
often distorted by artifacts associated with eye
movements, facial expressions, heart action, tightening the
neck muscles and others [13]. In this way, not only
desirable EEG signals (including SSVEP) are registered
but also noise signals and artifacts, which significantly
limit the possibility of effective SSVEP detection.
Therefore, an approach which enables the elimination of
artifacts from the EEG signal appears reasonable. To
eliminate such disturbances, frequency filters, spatial
filters, as well as Independent Component Analysis (ICA)
are the most commonly used [14],[15].
In the article we propose a system for automatic
SSVEP detection (Automatic Independent Component
Analysis - A-ICA). The method is based on an ICA
algorithm enhanced by automatic search of the component
associated only with SSVEP. With this approach, in an
indirect way, we eliminate the influence of artifacts and
noise usually associated with EEG signal registration. The
solution was tested using EEG signals coming from 5
users and registered for 4 stimuli of very close
frequencies: 5, 6, 7, 8Hz. The developed solution allowed
us to create a new procedure, which enables the SSVEP
detection. In order to choose the right ICA algorithm to
detect SSVEP we compared the most commonly used
ones.
II. MATERIALS
A database of EEG signals available on-line was used
in our experiments (Dataset BCI EEG SSVEP for four
classes of stimuli) [13]. EEG signals were registered for
five users, at the age of 23, 25 31, 42, and 46. Users sat
comfortably in a chair. A green LED of a 1cm diameter
was placed at a distance of about 1 meter from the eyes of
a user. EEG signals were recorded using g.USBAmp with
16 active electrodes. The electrodes were placed
according to the international 10-20 system at positions:
O2, AF3, AF4, P4, P3, F4, Fz, F3, FCz, Pz, C4, C3, CPz,
Cz, Oz, O1. EEG sampling frequency was 256Hz. The
signals were recorded using a Butterworth bandpass filter
(0.1-100Hz) and notch filter (48-52Hz) to correct a
technical artifact from the power network. Users were
stimulated with flickering LED light of frequencies: 5Hz,
6Hz, 7Hz and 8Hz. The stimulation lasted 30 seconds. All
sessions took place at the same time of the day to avoid
circadian influences on the measurements.
III. METHODS
The ICA enables the reconstruction of the source
signal S from a mixture of signals Xmk, where m depicts
the number of observed channels (rows of the matrix), and
k - the number of recorded samples (columns of the
matrix) [16]. The ICA problem comes down to finding
independent components Ymk. To sum up, we are looking
for the mixing matrix Amm and the separation matrix
W= which satisfy the equation:
X=AS (1)
The components of Y matrix correspond to the source
signals S:
Y=WX (2)
Many algorithms of finding the separation matrix W
has been proposed. Most often, these algorithms are based
on the search of the most "independent" to each other
components of Y matrix. Because there are many
possibilities of creating the Y matrix, the quality of the
ICA decomposition, to a large extent, depends on the used
algorithm. We also assume separation of other "source"
components associated with many activities as eye
movements - EOG, straining muscles - EMG, heart rate -
ECG.
The second important problem is an automatic
separation of interesting components that corresponds to
specific electrophysiological activity [16]. For this reason,
the ICA algorithms are typically used in off-line analysis -
we rely on visual assessment of the components by an
expert. The expert can also analyze weights assigned to
components in W matrix (commonly visualized on the
surface of the head) [17],[18]. This is time consuming and
difficult task. Several methods of automatic ICA
components detection are known from the literature
[19],[20]. For this purpose, features such as skewness and
kurtosis are used. However, we did not find in the
literature a solution for automatic ICA components
detection related to SSVEP.
We tested 23 commonly used ICA algorithms in order
to find the best method to determine the SSVEP
component. Detailed descriptions of the algorithms go far
beyond the scope of this article. The descriptions can be
found in the literature given by the name of each method
197
presented in Table I. We used ICA-LAB Toolbox [21] to
determine the ICA components.
For EEG signal recorded with N electrodes we can
obtain at most N independent components. Therefore, the
number of electrodes used during the recording is very
important. In addition, the obtained components are
disordered and we do not know their real amplitudes (and
the same the amplitudes of source signals). For the
selection of the proper component we can use only its
shape and the weights by which it was created. This is an
important constraint. In our experiments ICA components
were calculated for EEG signal taken from all sessions.
TABLE I. TESTED ICA METHODS
Algorithm The name of the algorithm / Publication
AMUSE - Algorithm for Multiple Unknown Source
ICA-1
Extraction based on EVD [22]
EVD2 (SOS BSS algorithm based on symmetric EVD)
ICA-2
[23]
ICA-3 SOBI - Second Order Blind Identification [24]
SOBI-RO - Robust SOBI with Robust
ICA-4
Orthogonalization [25]
SOBI-BPF - Robust SOBI with bank of Band-Pass
ICA-5
Filters [26]
WASOBI Blind signal separation by combining two
ICA-6
ICA algorithms: HOS-based EFICA [27]
EWASOBI Asymptotically Optimal Blind Separation
ICA-7
of Parametric Gaussian Sources [28]
ICA-8
FJADE (Flexible Joint Approximate Diagonalization
of Quadricovariance Matrices) [21]
JADE-op - Robust Joint Approximate Diagonalization
ICA-9 of Eigen Matrices (with optimized numerical
procedures) [29]
JADETD (HOS Joint Approximate Diagonalization of
ICA-10
Eigen matrices with Time Delays) [30]
QJADE Quadratic Join Aproximate Diagonalization
ICA-11
of Cumulant Matrices [21]
FJADE (Flexible Joint Approximate Diagonalization
ICA-12
of Quadricovariance Matrices) [21]
ICA-13 SAD (Sequential Approximate Diagonalization) [31]
ICA-14 FPICA (Fixed-Point ICA) [32]
ICA-15 POWERICA (Power iteration for ICA) [33]
ICA-16 EFICA (Efficient Variant of FastICA) [34]
ICA-17 COMBI (Combination of WASOBI and EFICA) [21]
MULTICOMBI (Multi-combination of WASOBI and
ICA-18
EFICA) [21]
SANG (Self Adaptive Natural Gradient algorithm with
ICA-19
nonholonomic constraints) [35]
ICA-20 NG-FICA (Natural Gradient - Flexible ICA) [36]
SIMBEC - SIMultaneous Blind Extraction using
ICA-21
Cumulants [37]
UNICA - Unbiased quasi Newton algorithm for ICA
ICA-22
[21]
ICA-23 Symmetric prewhitening[21]
A. Proposed A-ICA method
We have proposed our own fully automated procedure
(A-ICA) for SSVEP classification, which consists of
several steps (Fig.1). At first we calculate ICA
components (ICA) using one of available algorithms from
Table I. One of the difficulties of automatic ICA analysis
is the selection of appropriate components [38]. In our
algorithm to select the correct SSVEP component, weight
distributions associated with corresponding components
were used (Weight distributions). We assumed that this is
a single component. For each component, it is possible to
indicate the weights that formed it [39],[40]. For this
purpose, an appropriate row of the W matrix should be
selected. The row contains weights of the component
created from signals in X matrix.
Figure 1. The steps of A-ICA method.
The normalized weights were used as a set of input data
to 1-NN classifier (Classification) [41]. Then the
classifier was trained to automatically detect the
component which is associated with SSVEP (Component
selection). Further signal processing was carried out for
the selected SSVEP component. Time window of w width
overlapping of v samples was slid along the signal. For
each window position signal features were calculated
using FFT (Feature extraction). Frequency lines in f and
2f, 3f harmonics were taken as features. The calculated
features were used for teaching and testing the K-NN
classifier. Data were divided on training and testing sets
using ten cross validation (10-CV) test [42]. In this way,
the detection of SSVEP was made (SSVEP detection).
IV. RESULTS AND DISCUSSION
At first we examine to what extent the ICA methods
specified in Table 1, enable to obtain the SSVEP related
component (Fig.2). With N=16 EEG channels, for each
198
ICA method 16 components were calculated (ICA 1...23).
The result was 368 components for each user (S1 to S5).
Next for each component features associated with
frequency f and the harmonics 2f, 3f were calculated
(Feature extraction). As a result, 12 features for each
second of the component were obtained. For each
component the accuracy of SSVEP recognition was
calculated (SSVEP selection). For teaching and testing the
LDA classifier and 10-CV test was used. The
classification accuracies for individual components for
each of the ICA-(1,2,...,23) method were calculated. In
this manner, the usefulness of the component (related to
the accuracy of classification) was determined.
Figure 2. The procedure of ICA methods selection.
A typical classification accuracies for ICA-1 method (S1
user) are shown in Fig. 3. For four stimuli classes: 5, 6, 7,
8Hz random classifier accuracy (for equinumerous sets of
examples) is 0.25. In Fig.3, one can observe a
considerably distinctive component (#14) for which the
classification accuracy was 0.84.
Figure 3. Typical classification accuracies for ICA-1 method for S1
user.
Classification accuracy for indicated above fourteen
component are satisfactory and demonstrates its
usefulness in the process of SSVEP detection. For this
component, the distribution of weights assigned to
individual electrodes, on the surface of the head is drawn
in Fig. 4. As expected, the greatest weight values can be
observed on the visual cortex (back of the head). Similar S4 0.64 16, 23 0.57
effects were observed for the other users and other ICA S5 0.89 4, 6, 20 0.71
methods. Individual components (one or two) stand out
from the others with highest classification accuracies. It is difficult to indicate one, best ICA method. The
results are largely dependent on the quality of the
recorded EEG signal. For further analysis and
comparisons ICA-6 method was selected (Table 1), with
the highest average classification accuracy.
The main element of our research was to test the
usefulness of the proposed A-ICA method for automatic
SSVEP stimuli recognition. For this purpose, we
compared the results obtained for the proposed algorithm
with the results obtained without EEG pre-processing
using one of O1, O2 or Oz electrodes. We used time
window of 256 samples width 128 samples overlap.
Three harmonics were taken into account. LDA classifier
Figure 4. Distribution of weights assigned to individual electrodes and 10-CV test were used. In tables IV and V there are
(component number 14, ICA-1, S1 user). presented the classification accuracies for electrodes O1,
O2 and Oz (without EEG pre-processing) and for A-ICA
The result of STFT analysis for this component is method for one and three harmonics.
presented in Fig. 5 in the form of spectrogram. Four
stimuli can be seen: 5Hz (0-30s), 6Hz (30-60s), 7Hz (60- TABLE IV. CLASSIFICATION ACCURACY FOR A-ICA METHOD AND FOR
O1, O2 AND OZ ELECTRODES (NO EEG SIGNAL PRE-PROCESSING) FOR
90s), 8Hz (90-120s). In the spectrogram, one can observe THREE HARMONICS
the presence of harmonics. It is worth noting the
appearance of the 4th harmonic for 5Hz and 6Hz stimuli. Subject O2 Oz O1 A-ICA
S1 0.46 0.6 0.62 0.88
S2 0.39 0.48 0.43 0.74
S3 0.39 0.36 0.39 0.53
S4 0.54 0.55 0.55 0.60
S5 0.42 0.58 0.62 0.89

TABLE V. CLASSIFICATION ACCURACY FOR A-ICA METHOD AND FOR

O1, O2 AND OZ ELECTRODES (NO EEG SIGNAL PRE-PROCESSING) FOR
THE FIRST HARMONIC

Subject O2 Oz O1 A-ICA
S1 0.47 0.59 0.53 0.8
S2 0.4 0.5 0.42 0.65
S3 0.3 0.4 0.36 0.47
S4 0.45 0.46 0.45 0.52
S5 0.47 0.56 0.59 0.82
Figure 5. Spectrogram of the component number 14 (ICA-1 method,
S1 user) for stimuli: 5, 6, 7, 8Hz.
The results presented in Tables IV and V allow us to
Four best ICA methods were indicated, taking into conclude that the developed method can effectively
consideration the average classification accuracy. They identify SSVP component. The average classification
were: ICA-6 average classification accuracy 0.73, ICA-3 accuracy for three harmonics, obtained with the use of the
average classification accuracy 0.72, ICA-4 average SSVEP component, increased for A-ICA algorithm by an
classification accuracy 0.70. The maximum values of average of 28%, 21% and 20% in comparison with the
classification accuracies for individual users for the best direct method comprising O2, Oz or O1 electrode. The
ICA method and the average values of classification average classification accuracy for the first harmonic for
accuracies for all ICA - (1,2,..,23) methods are presented A-ICA algorithm increased by an average of 24%, 15%
in Table III. and 18% in comparison with the direct method
comprising O2, Oz, O1 electrodes.
TABLE III. THE MAXIMUM VALUES OF THE CLASSIFICATION ACCURACY
WITH THE USE OF THE SELECTED COMPONENT, FOR SELECTED ICA In Table VI there is a comparison of the average
METHODS
classification accuracies obtained for each ICA method.
Best
Numbers of the Average The highest classification accuracies are bolded. The
Subject best ICA accuracy for all results are difficult to compare with other reported in the
accuracy
methods ICA methods
S1 0.90 21 0.77
literature. This is due to the fact that the quality of
S2 0.75 8 0.62 recorded EEG signal, signal processing methods and the
S3 0.60 13 0.51

199
method of classifier testing have significant impact on the [4] M. Higger, M. Akcakaya, H. Nezamfar, G. LaMountain, U.
Orhan, and D. Erdogmus, A Bayesian Framework for Intent
obtained classification accuracy.
Detection and Stimulation Selection in SSVEP BCIs, IEEE
Signal Process. Lett., vol. 22, no. 6, pp. 743747, Jun. 2015.
TABLE VI. CLASSIFICATION ACCURACY FOR ALL ICA METHODS AND [5] A. Materka and P. Poryzala, High-speed noninvasive brain-
FOR THREE HARMONICS
computer interfaces, in 2013 6th International Conference on
Classification Classification Human System Interactions (HSI), 2013, pp. 712.
Classification [6] B. Allison, T. Luth, D. Valbuena, A. Teymourian, I. Volosyak,
accuracy accuracy
Algorithm accuracy and A. Graser, BCI Demographics: How Many (and What
I, II I,II,III
I harmonic Kinds of) People Can Use an SSVEP BCI?, IEEE Trans.
harmonics harmonics
ICA-1 0.57 0.61 0.62 Neural Syst. Rehabil. Eng., vol. 18, no. 2, pp. 107116, Apr.
ICA-2 0.54 0.57 0.58 2010.
[7] Y. Wang, Y.-T. Wang, and T.-P. Jung, Visual stimulus design
ICA-3 0.65 0.68 0.70
for high-rate SSVEP BCI, Electron. Lett., vol. 46, no. 15, pp.
ICA-4 0.66 0.71 0.72
10571058, Jul. 2010.
ICA-5 0.60 0.62 0.63 [8] P.-L. Lee, C.-L. Yeh, J. Y.-S. Cheng, C.-Y. Yang, and G.-Y.
ICA-6 0.66 0.71 0.72 Lan, An SSVEP-Based BCI Using High Duty-Cycle Visual
ICA-7 0.62 0.64 0.65 Flicker, IEEE Trans. Biomed. Eng., vol. 58, no. 12, pp. 3350
ICA-8 0.56 0.61 0.62 3359, Dec. 2011.
ICA-9 0.54 0.57 0.59 [9] E. Kalunga, K. Djouani, Y. Hamam, S. Chevallier, and E.
ICA-10 0.56 0.61 0.61 Monacelli, SSVEP enhancement based on Canonical
ICA-11 0.60 0.64 0.64 Correlation Analysis to improve BCI performances, in
ICA-12 0.65 0.68 0.69 AFRICON, 2013, 2013, pp. 15.
ICA-13 0.65 0.69 0.69 [10] L.-W. Ko, S.-C. Lin, M.-S. Song, and O. Komarov,
ICA-14 0.51 0.53 0.54 Developing a few-channel hybrid BCI system by using motor
ICA-15 0.51 0.57 0.58 imagery with SSVEP assist, in 2014 International Joint
ICA-16 0.59 0.64 0.65 Conference on Neural Networks (IJCNN), 2014, pp. 41144120.
ICA-17 0.61 0.63 0.64 [11] L. Maggi, S. Parini, L. Piccini, G. Panfili, and G. Andreoni, A
ICA-18 0.62 0.66 0.66 four command BCI system based on the SSVEP protocol, in
ICA-19 0.54 0.55 0.58 28th Annual International Conference of the IEEE Engineering
ICA-20 0.62 0.68 0.68 in Medicine and Biology Society, 2006. EMBS 06, 2006, pp.
12641267.
ICA-21 0.62 0.66 0.67
[12] S. Karan, A literature survey on the contemporary
ICA-22 0.51 0.54 0.54
methodologies used in brain computer interface for spelling
ICA-23 0.58 0.64 0.64 application, in 2013 International Conference on Human
Computer Interactions (ICHCI), 2013, pp. 15.
[13] M. Kolodziej, A. Majkowski, and R. J. Rak, A new method of
V. CONCLUSIONS spatial filters design for brain-computer interface based on
steady state visually evoked potentials, in 2015 IEEE 8th
We tested more than 20 ICA algorithms in the task of
International Conference on Intelligent Data Acquisition and
EEG signal decomposition for SSVEP component Advanced Computing Systems: Technology and Applications
recognition. The best results were obtained for algorithms: (IDAACS), 2015, vol. 2, pp. 697700.
WASOBI Blind signal separation (ICA-6 - average [14] M. R. H. Samadi and N. Cooke, VOG-enhanced ICA for
SSVEP response detection from consumer-grade EEG, in
classification accuracy 0.73) [27], SOBI - Second Order
Signal Processing Conference (EUSIPCO), 2014 Proceedings
Blind Identification (ICA-3 - average classification of the 22nd European, 2014, pp. 20252029.
accuracy 0.72) [24] and SOBI-RO - Robust SOBI with [15] X. Ruan, K. Xue, and M. Li, Feature extraction of SSVEP-
Robust Orthogonalization (ICA-4 - average classification based brain-computer interface with ICA and HHT method, in
2014 11th World Congress on Intelligent Control and
accuracy 0.70) [25]. The proposed A-ICA algorithm
Automation (WCICA), 2014, pp. 24182423.
enables automatic classification of EEG signals taking [16] C. A. Joyce, I. F. Gorodnitsky, and M. Kutas, Automatic
into consideration SSVEP. Classification results were removal of eye movement and blink artifacts from EEG data
about 20% better than for the direct analysis of the EEG using blind component separation, Psychophysiology, vol. 41,
no. 2, pp. 313325, Mar. 2004.
signal, registered with one electrode placed on visual [17] S. Makeig, S. Enghoff, T.-P. Jung, and T. J. Sejnowski, A
cortex (O1, O2 or Oz). The algorithm is fully automatic in natural basis for efficient brain-actuated control, IEEE Trans.
off-line mode. It is worth considering to adapt it to work Rehabil. Eng., vol. 8, no. 2, pp. 208211, Jun. 2000.
in real time, for the use in brain-computer interfaces. [18] M. P. Guimaraes, D. K. Wong, E. T. Uy, L. Grosenick, and P.
Suppes, Single-Trial Classification of MEG Recordings, IEEE
REFERENCES Trans. Biomed. Eng., vol. 54, no. 3, pp. 436443, Mar. 2007.
[19] D. E. Kusumandari, H. Fakhrurroja, A. Turnip, S. S.
[1] R. J. Rak, M. Koodziej, and A. Majkowski, Brain-computer Hutagalung, B. Kumbara, and J. Simarmata, Removal of EOG
interface as measurement and control system the review paper, artifacts: Comparison of ICA algorithm from recording EEG,
Metrol. Meas. Syst., vol. 19, no. 3, pp. 427444, 2012. in 2014 2nd International Conference on Technology,
[2] Y.-J. Chen, A. R. A. See, and S.-C. Chen, SSVEP-based BCI Informatics, Management, Engineering, and Environment
classification using power cepstrum analysis, Electron. Lett., (TIME-E), 2014, pp. 335339.
vol. 50, no. 10, pp. 735737, May 2014. [20] N. S. Cheok and P. Raveendran, Removal of EOG Artifacts
[3] C.-H. Han, H.-J. Hwang, and C.-H. Im, Classification of visual Using ICA Regression Method, in 4th Kuala Lumpur
stimuli with different spatial patterns for single-frequency, International Conference on Biomedical Engineering 2008, D.
multi-class SSVEP BCI, Electron. Lett., vol. 49, no. 22, pp. N. A. A. Osman, D. F. Ibrahim, D. W. A. B. W. Abas, H. S. A.
13741376, Oct. 2013. Rahman, and D. H.-N. Ting, Eds. Springer Berlin Heidelberg,
2008, pp. 226229.

200
[21] A. Cichocki, S. Amari, K. Siwek, T. Tanaka, A. H. Phan, and
others, ICALAB Toolboxes h http://www. bsp. brain. riken.
jp, ICALAB I, 2004.
[22] L. Tong, R. Liu, V. C. Soon, and Y.-F. Huang, Indeterminacy
and identifiability of blind identification, IEEE Trans. Circuits
Syst., vol. 38, no. 5, pp. 499509, May 1991.
[23] P. Georgiev and A. Cichocki, Blind source separation via
symmetric eigenvalue decomposition, in Signal Processing and
its Applications, Sixth International, Symposium on. 2001, 2001,
vol. 1, pp. 1720 vol.1.
[24] A. Belouchrani, K. Abed-meraim, J. F. Cardoso, and E.
Moulines, Second Order Blind Separation of Temporally
Correlated Sources. 1993.
[25] A. Belouchrani and A. Cichocki, Robust whitening procedure
in blind source separation context, Electron. Lett., vol. 36, no.
24, pp. 20502051, Nov. 2000.
[26] R. R. Gharieb and A. Cichocki, Second-order statistics based
blind source separation using a bank of subband filters, Digit.
Signal Process., vol. 13, no. 2, pp. 252274, Apr. 2003.
[27] P. Tichavsky, Z. Koldovsky, E. Doron, A. Yeredor, and G.
Gomez-Herrero, Blind signal separation by combining two
ICA algorithms: HOS-based EFICA and time structure-based
WASOBI, in Signal Processing Conference, 2006 14th
European, 2006, pp. 15.
[28] E. Doron and A. Yeredor, Asymptotically Optimal Blind
Separation of Parametric Gaussian Sources, in Independent
Component Analysis and Blind Signal Separation, C. G.
Puntonet and A. Prieto, Eds. Springer Berlin Heidelberg, 2004,
pp. 390397.
[29] J.-F. Cardoso and A. Souloumiac, Blind Beamforming for Non
Gaussian Signals, IEE Proc.-F, vol. 140, pp. 362370, 1993.
[30] P. Georgiev and A. Cichocki, Robust Blind Source Separation
Utilizing Second and Fourth Order Statistics, in Artificial
Neural Networks ICANN 2002, J. R. Dorronsoro, Ed.
Springer Berlin Heidelberg, 2002, pp. 11621167.
[31] X.-L. Li and X.-D. Zhang, Nonorthogonal Joint
Diagonalization Free of Degenerate Solution, IEEE Trans.
Signal Process., vol. 55, no. 5, pp. 18031814, May 2007.
201
[32] A. Hyvrinen and E. Oja, A Fast Fixed-Point Algorithm for
Independent Component Analysis, Neural Comput., vol. 9, no.
7, pp. 14831492, Oct. 1997.
[33] S. Ding, Independent Component Analysis without
Predetermined Learning Parameters, in The Sixth IEEE
International Conference on Computer and Information
Technology, 2006. CIT 06, 2006, pp. 135135.
[34] Z. Koldovsky, P. Tichavsky, and E. Oja, Efficient Variant of
Algorithm FastICA for Independent Component Analysis
Attaining the Cram amp;#201;r-Rao Lower Bound, IEEE
Trans. Neural Netw., vol. 17, no. 5, pp. 12651277, Sep. 2006.
[35] S. Amari, T. P. Chen, and A. Cichocki, Nonholonomic
orthogonal learning algorithms for blind source separation,
Neural Comput., vol. 12, no. 6, pp. 14631484, Jun. 2000.
[36] S. Choi, A. Cichocki, and S.-I. Amari, Flexible Independent
Component Analysis, J. VLSI Signal Process. Syst. Signal
Image Video Technol., vol. 26, no. 12, pp. 2538, Aug. 2000.
[37] S. Amari, Natural Gradient Learning for Over- and Under-
Complete Bases in ICA, Neural Comput., vol. 11, no. 8, pp.
18751883, Nov. 1999.
[38] N. Mammone, F. La Foresta, and F. C. Morabito, Automatic
Artifact Rejection From Multichannel Scalp EEG by Wavelet
ICA, IEEE Sens. J., vol. 12, no. 3, pp. 533542, Mar. 2012.
[39] T.-P. Jung, S. Makeig, C. Humphries, T.-W. Lee, M. J.
McKEOWN, V. Iragui, and T. J. Sejnowski, Removing
electroencephalographic artifacts by blind source separation,
Psychophysiology, vol. 37, no. 2, pp. 163178, Mar. 2000.
[40] T.-P. Jung, C. Humphries, T. Lee, S. Makeig, M. J. Mckeown,
V. Iragui, and T. J. Sejnowski, Extended ICA Removes Artifacts
from Electroencephalographic Recordings. 1997.
[41] E. Bax, Validation of -Nearest Neighbor Classifiers, IEEE
Trans. Inf. Theory, vol. 58, no. 5, pp. 32253234, May 2012.
[42] K. Yang, H. Wang, G. Dai, S. Hu, Y. Zhang, and J. Xu,
Determining the repeat number of cross-validation, in 2011
4th International Conference on Biomedical Engineering and
Informatics (BMEI), 2011, vol. 3, pp. 17061710.