Complex ICA-R

Jagath C. Rajapakse, Senior Member, IEEE, and Wenda Chen

Abstract— The Complex Independent Component Analysis weights [15]. This significantly improved the efficiency of
(CICA) which extends Independent Component Analysis (ICA) ICA-R. The application of ICA-R have shown promise in
to complex signals has found applications in various fields. The several applications such as EEG [16][17] and functional
ICA with Reference (ICA-R) has recently gained popularity
in semi-blind separation of signals when a priori information MRI [12].
of the desired sources are available in the form of reference In this paper, we introduce Complex ICA with Reference
signals. This paper extends the framework of ICA-R to complex (CICA-R) for extraction of desired complex source signals
signals and demonstrates the use of Complex ICA-R (CICA-R) from their mixtures. An approach to CICA-R, making use of
with applications to both synthetic data and real speech data.
the information of only the signal magnitude in the contrast
Our experiments indicate that CICA-R is more effective than
ICA, ICA-R, or CICA, in separation of complex signals when function has been earlier proposed [18]. The method is thus
reference signals relating to source signals are available. capable of dealing with only circular signals. In order to cater
CICA-R for a wider range of signals, we will extend the Fast
I. I NTRODUCTION CICA framework with a different negentropy measurement
Independent Component Analysis (ICA) has become a as the contrast function. The learning is performed by using
popular method for blind signal separation, which is based a Quasi-Newton method.
on the assumption that original sources are statistically This paper is organized as follows: Section 2 gives a
independent [1][2][3]. Complex ICA (CICA) was introduced brief introduction to ICA, CICA, and ICA-R; Section 3
to deal with complex signals [4] and has found applications introduces the framework of CICA-R; in Section 4, we
in digital signal communication [5], radar communication give a quantitative evaluation of the proposed method using
[6], and functional brain imaging such as fMRI and EEG synthetic data and compare its performance with earlier
[7]. methods; further, applications of CICA-R on speech data
The Fast ICA algorithm [3], a popular and efficient method and time domain convolutive mixtures are demonstrated; and
of performing ICA in time domain, has been extended Section 5 concludes the paper with a discussion.
to complex domain. As the non-linear function generating
higher-order statistics in the contrast function is evaluated II. I NDEPENDENT C OMPONENT A NALYSIS (ICA)
on the magnitude of the signals, Fast CICA is able to
We consider time-varying signals that are linearly and
separate only circularly symmetric sources [4][8]. The com-
instantaneously mixed. Consider a set of N mixed sig-
plex maximization of non-Gaussianity (CMN) algorithm has
nals, x(t) = (x1 (t), x2 (t), . . . xN (t))T and source signals,
been introduced to handle both symmetric and asymmetric
s(t) = (s1 (t), s2 (t), . . . sN (t))T where t denotes the time.
sources by incorporating both the magnitude and the phase
For instantaneous mixtures of signals and kth input signal is
information in the non-linear function [9]. The method of
given by
joint approximate diagonalization of eigenmatrices (JADE)
N
has been extended to complex domain, which is suitable only X
xk (t) = aki si (t) (1)
for a small number of sources [10]. The strong-uncorrelating
i=1
transform (SUT) is an efficient way to perform CICA but
requires data to be strictly non-circular with distinct spectra where A = {aki }N ×N is the linear mixing matrix and aki
of the pseudo-covariance matrices [8]. The above CICA are assumed to be real-valued coefficients.
algorithms make certain assumptions on the properties of Assuming independence of sources, the ICA estimates the
data and are unable to use any a priori knowledge of sources demixing matrix to find the original sources. If the output
when available. of the ICA is y(t) = (y1 (t), y2 (t), . . . , yN (t))T , then y(t) =
ICA with Reference (ICA-R) [11][12], a special case of W x(t) where W denotes the desired demixing matrix. In this
constrained ICA (cICA) [13][14], provides a more accurate work, we assume complete ICA where the number of sources
and efficient way of extracting desired sources than ICA is equal to the number of input mixtures and the mixture is
when a priori information about the sources are available noise free. The methods could be extended when the mixtures
in the form of reference signals. The corresponding ref- are under-complete or over-complete, and corrupted by noise,
erence signals having some characteristics of the desired investigations into which are out of the scope of the present
sources guide the signal separation. The closenesses between paper. In what follows, we often omit time variable t in our
the output signals and reference signals are formulated as formula for simplicity.
constraints to the contrast function in ICA-R [12]. Lin et Different contrast functions are used in ICA algorithms
al. proposed a fast one-unit ICA-R where the uncorrelation to maximize the independence among the output signals in
equality constraint of ICA-R is handled by normalizing the order to estimate the demixing matrix. Negentropy J is a

978-1-4244-8126-2/10/$26.00 ©2010 IEEE

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on January 21,2025 at 07:31:25 UTC from IEEE Xplore. Restrictions apply.
often used as the contrast function to measure independence Since the non-linear function G is only evaluated on the
among the extracted sources: magnitude of the output signal, it is capable of separating
only the circular symmetric signals. The Fast CICA algo-
J(y) = H(ygauss ) − H(y) rithm can be used only in applications where signals are
where ygauss is a Gaussian random variable with the same circularly symmetric such as EEG. For circularly symmetric
variance as the signal y and H denotes the entropy of signals x, the covariance matrix is rotationally invariant:
the signal. Because of its accuracy and precision, here, we E{(x − E{x})(x − E{x})T } = 0.
consider negentropy as the contrast function and the Fast The CMN algorithm is able to handle both circular and
ICA algorithm for separating the source components [3]. non-circular signals [9]. The contrast function is based on the
The Fast ICA achieves its efficiency by using the following non-Gaussianity function in which the non-linear function is
approximation for negentropy: evaluated on both the phase and the magnitude of the sources:

J(y) = E{G(wT x)}. (2) J(w) = E{|G(wH x)|2 } (5)

By properly selecting the nonlinear function G for ap- where w is the demixing weight vector.
proximating non-Gaussianity, higher-order statistics of the B. ICA with Reference (ICA-R)
signals are generated so that the fast-ICA can effectively and
ICA with Reference (ICA-R) was proposed to extract the
efficiently separate the input signals spanning over a wide
desired signals when a priori knowledge of the sources
range of density functions [3].
is available in the form of reference signals [13][14]. The
A. Complex ICA (CICA) reference signals are closer to the desired signals and share
some characteristics of the desired signals such as the period,
The classical ICA is capable of separating linearly mixed
zero crossings, peaks, etc. They are incorporated into the
source signals in time-domain. One particular case of Com-
contrast function as constraints to the extracting signals in
plex ICA is when the real signals are convolutively mixed.
the optimization. As ICA-R extracts only the desired sources,
For mixtures, where the source signals are convolutively
the indeterminacy problem does not arise. It converges faster
mixed [7][19], the k th input signal can be written as
and is more accurate than the classical ICA when reference
N X
X R signals are available [12].
xk (t) = aki (τ )si (t − τ ), for k = 1, 2, . . . N (3) Suppose we have l number of reference signals r =
i=1 τ =1 [r1 , r2 , . . . , rl ] corresponding to l source signals. The ICA-R
where aki (τ ) denotes the convolution filter of source i to the algorithm is achieved by modifying the ICA contrast function
input sensor k. Taking the Fourier Transform of (3), by adding closeness measures as inequality constraints:
N
X max C(y) = Σli=0 J(yi )
yi
xk (f ) = Aki (f )si (f ) (4)
i=1 subject to ε(yi , ri ) − ξi ≤ 0, ∀i = 1, 2, . . . l (6)
PR τ −1
where Aki (f ) = τ =1 aki (τ )e−j2πf R and R is the length The closeness between the estimated output yi and the cor-
of the filter’s impulse response. That is, the linear mixing responding reference ri is measured by some norm ε(yi , ri ).
matrix A is complex for convolutive mixtures of real signals The output with the minimum value of ε(yi , ri ) is the desired
in time domain. independent component (IC). That is, ε(yi∗ , ri ) < ε(yio , ri )
The indeterminacy of permutation and scaling is inherent where yi∗ denotes the IC closest to the reference ri and
to ICA [20] and arises in the reconstruction of sources yio denotes the output having the next closest value. The
in CICA [21]. This could be ameliorated by using short ξi in the constraint is given by ξi = [ε(yi∗ , ri ), ε(yio , ri )].
time Fourier transform (STFT) which divides the frequency In the original formulation of ICA-R, constraints to avoid
spectrum into smaller frequency bands [22]. Then, x(f, t) = correlation among ICs were also used but such constraints
A(f )s(f, t) and the demixing matrix W (f ) is derived using can be removed by pre-whitening the data and normalizing
signals in each frequency band: y(f, t) = W (f )x(f, t). The the weight at each iteration [15].
signal is reconstructed by using the inverse STFT. By using III. C OMPLEX ICA-R (CICA-R)
short bands of signals in STFT, the permutation problem can
be circumvented to a certain extent. However, if the signals In this section, we will introduce the framework of CICA-
are of large bandwidths, the permutation problem persists R (CICA-R): first, one-unit CICA-R for extracting a single
because the desired signal components from wider bins have source of the mixture and second, the multi-unit case when
to be identified and compiled. multiple sources are to be extracted. The demixing matrix W
In CICA, source signals s and the mixed signals x are is obtained by maximizing the contrast function by using an
complex. The Fast CICA algorithm uses the negentropy as augmented Lagrange multipliers method [14]. The complex
the the contrast function [8] as follows: input signals x are assumed to be pre-whitened such that
E{xxH } = I where H denotes the Hermitian transpose.
J(w) = E{G(|wH x|2 )} In CICA-R, permutation and indeterminacy problems of

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on January 21,2025 at 07:31:25 UTC from IEEE Xplore. Restrictions apply.
the output signals does not exist as the reference signals When choosing certain classes of nonlinear functions, (HJ −
constrains the converging targets. µI) can be real-valued and hence be removed from the
learning equation. So, the rules of weight updating become
A. One-unit CICA-R
In one-unit CICA-R, extracting only a single source is wnew = −E{xG∗ (y)g(y)} + µE{x(wH x − r)∗ }
0 0
interested from input mixtures of sources by using one −µwn + E{G (y)G ∗ (y)}w
00
reference signal. We use the contrast function introduced in +E{xxT G∗ (y)G (y)}w∗ (14)
(5).
The closeness between the output signal y and the refer- where we have constraint ||w|| = 1 replaced by normalizing
wnew
ence signal r is measured by the mean square error ε(y, r) wnew = ||w new || .

in complex domain. That is, g(w) = ε(y, r) − ξ ≤ 0 where The update rule of µ is as follows
y = wH x. In vector format: µnew = max{µ + γg(w), 0} (15)
1
g(w) = (y − r)(y − r)H − ξ For one-unit CICA-R, unlike the original ICA-R, the decor-
N
1 H relation step is avoided and thereby the distortions of the
= (w x − r)(wH x − r)H − ξ (7) recovered sources are avoided and the algorithm becomes
N
computationally more efficient.
where ξ is gradually increased from a small initial value
to a larger value to ensure proper convergence to the B. Multi-unit CICA-R
desired independent component [12]. Since the range When a number of source signals of interests are mixed
of ξ ∈ [min{ε(y, r)}, ε(yi , ri )) where min{ε(y, r)} = in the inputs and the corresponding reference signals are
mini (ε(yi , ri )), the value of ξ is initially fixed and increased all available, the above one-unit CICA-R algorithm can
gradually. y and r are normalized to have zero means and be generalized to multi-unit CICA-R. All the optimization
unit variances. operations are performed on each element of the contrast
The Lagrange function incorporating the equality con- function in vector form, simultaneously and independently.
straint f (w) = E{|wH x|2 } − 1 = ||w||2 − 1 = 0 and So, we maximize
inequality constraints: l
X
H 1 2 J(wi ) subject to g(W ) ≤ 0 (16)
L1 (w, µ) = E{|G(w x)| } − [max{µ + γg(w), 0}2
2γ i=1
1 where the demixing matrix, W = (w1 , w2 , · · · wl )T and
−µ2 ] − λ(f (w)) − γ||f (w)||2 (8)
2 g(W ) = (g(w1 ), g(w2 ), · · · g(wl ))T . The inequality con-
is reduced to [4][15]: straints in vector format is given by
L1 (w, µ) = J(w) − h(w) 1
g(wi ) = (yi − r)(yi − r)H − ξ
1 N
= E{|G(wH x)|2 } − [max{µ + γg(w), 0}2 1 H
2γ = (w x − r)(wiH x − r)H − ξi (17)
−µ2 ] (9) N i
The corresponding Lagrangian function[4] L2 is:
1
where h(w) = 2γ [max{µ + γg(w), 0}2 − µ2 ]. l
The Quasi-Newton learning rules are given by X
L2 (W, µ) = (E{|G(wiH x)|2 }
new ∇L1 (w, µ) i=1
w =w− 2 (10) 1
∇ L1 (w, µ) − [max{µi + γi g(wi ), 0}2 − µ2i ])(18)
where 2γi
where µ = (µ1 , µ2 , · · · µl )T , γ = (γ1 , γ2 · · · γl )T .
∇L1 (w, µ) = ∇J(w) − ∇h(w)
The weight updating is given by
= E{xG∗ (y)g(y)} − µE{x(wH x − r)(11)
∗
}
∇L2 (W, µ)
2
∇ L1 (w, µ) = HJ − µI (12) W new = W − (19)
∇2 L2 (W, µ)
in which HJ is the complex Hessian HJ = ∇2w J(w) and W+
and followed by the normalization, W new = ||W + || . It can
µI = ∇2w h(w).
be shown that
Substituting the equations above and HJ w =
0 0 00
E{G (y)G ∗ (y)}w + E{xxT G∗ (y)G (y)}w∗ as in W+ = −E{xG∗ (y)g(y)} + µT E{x(W H x − r)∗ }
0 0
[9] into equation (11), we have: −µT W + E{G (y)G ∗ (y)}W
00
(HJ − µI)wnew = −E{xG∗ (y)g(y)} +E{xxT G∗ (y)G (y)}W ∗ (20)
0 0
H ∗ ∗
+µE{x(w x − r) } − µw + E{G (y)G (y)}w At the matrix level, we follow the symmetric orthogonaliza-
00 1
+E{xxT G∗ (y)G (y)}w∗ (13) tion as the matrix de-correlation step W = W (W H W )− 2 to

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on January 21,2025 at 07:31:25 UTC from IEEE Xplore. Restrictions apply.
update the weights [3]. The Lagrange multipliers are updated 1) All sources are circular: one complex modulated chirp
as signals whose frequency change with time, one 4-
QAM (Quadrature Phase Shift Keying) signal, and
µnew = max{µ + γ T g(W ), 0} (21)
five complex signals, sj = rj (cos φj + i sin φj ), with
IV. E XPERIMENTS AND R ESULTS magnitudes r each of which is distributed as binomial,
gamma, Poisson, hypergeometric, and beta densities,
We demonstrate the use of CICA-R by using synthetic
respectively, and the phase φ is uniformly distributed.
datasets as well as real speech signals. The reference signals
2) Five non-circular signals and two circular signals (chirp
r representing the prior knowledge of the complex signals
and 4-QAM): five randomly distributed signals: the real
were generated using the following approach: assuming that
component, real(si ), and imaginary component, im(si ),
the knowledge of the frequency of the signal is available,
of the random signal si have the same distributions,
the sign function was applied on both the real parts and
binomial, gamma, Poisson, hypergeometric, and beta,
imaginary parts of the complex signals.
respectively, but different variances so that they are
The performance of the CICA-R algorithm will be com-
non-circularly distributed [23].
pared with ICA, Fast CICA [4], and CMN algorithm [9].
For CICA-R and CMN algorithms, the non-linear function In all the experiments, each dataset was generated 25
G(z) = z 1.25 was used as it has performed well earlier [9]. times by using a random mixing matrices and the mean
The quality of signal extraction is measured by the signal- (± std. dev.) of PI values is reported. The performance on
to-noise-ratio (SNR) values. The SNR value of the extracted extraction of sources were compared with Fast CICA and
signal yi is given by CMN algorithms. The values of performance indicies (PI)
are are shown in Table 1. As seen, the CICA-R showed
||si − s¯i || superior performance on the separation of sources in all
SNR = 20 log10
||si − yi || types of synthetic datasets. It shows that the current analytic
where si is the corresponding source signal and s¯i is its G function is good enough for separating signal mixtures
mean. As a measure of quality of separation of synthetic data, comprising of sources having a wide range of statistical
we find a normalized performance index (PI) to measure the properties, and the use of reference signals enhances the
quality of source separation [14]: separation.
 
l
! l B. Convoluted Mixtures
1 X 1 X
PI = rPIi +  cPIj  For convolutive mixtures of signals, we used STFT and
2l(l − 1) i=1 2l j=1
separated the frequency components at different frequency
l bands. The original signals comprised of one simulated EEG
X |pij |2 pulse signal (s1 ), one bandpass periodic signal mixed by
where rPIi = − 1 in which pij denotes the
j=1
(maxk |pik |)2 sinusoids (s2 ), one sinusoid signal (s3 ), and one periodic
(i, j) th element of the permutation matrix P = W H A, waveform signal (s4 ). The mixed signals were generated
Xl
|pij |2 by using convolution filters where the convolution matrix
and cPIj = − 1 in which pij denotes the A = {aij }4×4 was applied on the original signals [21]. The
(maxk |pkj |)2
i=1 convolution matrix was randomly generated. The sampling
elements in P corresponding to the j th IC in the desired
rate is 160Hz and each frequency band size is 5Hz. In STFT,
subset of sources.
we use Hanning windows with the length covering one period
A. Synthetic Datasets of the desired signals and an overlap of 50%. The overlap
In order to obtain a quantitative assessment of our algo- for signal reconstruction is 50%.
rithm, we combine complex signals of different properties The extracted signals by ICA-R, Fast CICA, CMN and
and then randomly mixed to obtain synthetic datasets. We CICA-R algorithms are shown in figure 1 and the corre-
consider the following source signals: sponding SNR values are given in Table 2. As indicated by
lower SNR, CICA-R showed superior performance over the
1) complex random signals, sj = rj (cos φj + i sin φj )
other methods. The average PI values in the frequency bins
where r denotes the magnitude and φ denotes the
for signals extracted from CICA-R, Fast CICA, and CMN
phase. Both the amplitude and phase could take dif-
algorithms were 3.63(±0.44), 4.18(±0.57), and 4.76(±0.82),
ferent forms of densities.
respectively.
2) A circular complex modulated chirp signal x(t) =
2
e(αt +βt+γ)i C. Cocktail Party Problem
3) A quadrature amplitude modulated signal (4-QAM) Speech signals are non-stationary and cover a wide band
The chirp and 4-QAM signals were generated to be of frequency. In a cocktail party problem, they are assumed
circularly symmetric. Two synthetic datasets were created to be linearly or partially convolutively mixed. Therefore,
by randomly mixing above source signals so that the per- CICA-R should be more appropriate in separating speech
formance of CICA-R can be evaluated on both circular and signals in a cocktail party problem [15]. In this section, we
non-circular sources. illustrate how CICA-R could be used to extract the speakers’

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on January 21,2025 at 07:31:25 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
converges faster, and extracts signals with higher accuracy.
T HE PERFORMANCE INDICES (PI) FOR EXTRACTION OF SIGNALS FROM
Compared to existing time-domain and complex-domain ICA
SYNTHETIC DATASETS BY FAST CICA, CMN, AND CICA-R
techniques, CICA-R satisfactorily deals with complex signals
ALGORITHMS
with specific distributions and is also good for mixtures of
data Fast CICA CMN CICA-R
d1 1.210±0.221 1.330±0.226 0.365±0.040 signals with different properties and distributions.
d2 1.194±0.162 1.146±0.201 0.347±0.089 Complex signals are found in many applications, for
example in convoluted mixtures of signals in time domain.
TABLE II Our synthetic datasets were designed to represent various cir-
SNR VALUES ( D B) OF EXTRACTED SIGNALS FROM CONVOLUTED cular properties of signals encountered in various application
MIXTURES BY ICA-R,FAST CICA, CMN, AND CICA-R ALGORITHMS domains such as wireless communications and modulations,
source ICA-R Fast CICA CMN CICA-R
biomedical signal processing such as EEG [7] and complex-
S1 7.33 ± 0.65 8.47 ± 0.94 8.08 ± 0.94 16.31 ± 1.02 valued fMRI [25], and data mining [3]. They also represent
S2 6.11 ± 1.07 7.44 ± 0.80 9.36 ± 1.07 14.17 ± 0.69 different circularity properties of the Fourier Transformed
S3 7.32 ± 0.49 8.29 ± 0.68 9.64 ± 0.58 13.51 ± 0.74 time domain signals [7]. The results with synthetic and real
S4 5.89 ± 1.03 6.09 ± 0.74 7.44 ± 0.69 12.15 ± 0.83
data indicate that CICA-R could be useful and accurate
method for separating signals in a wide range of applications.
Compared to CICA, the CICA-R does not suffer from the
voice in a cocktail party. Since the Fourier Transform has the indeterminacy problem of scale and order.
linear property, the extraction of sources are performed in We demonstrated the efficacy of CICA in separating
the frequency domain. Therefore, for linear mixing matrix, speech signals in a real cocktail party problem. The speech
we can find the same demixing matrix for all the frequency signals may be partially convolved and delayed, so the
bands. We then combine the frequency components to the complex-domain methods are more appropriate. In addition,
same sequence of complex numbers. There will be no the CICA-R is particularly good because the indeterminacy
permutation or scaling problems when dealing purely with problem is eliminated in the frequency domain with the
complex signals in the frequency domain [21]. help of reference signals. The reference signals of speech
The original sample speech signals dataset obtained from signals are estimated by their pitch and periodicity, which
BSS DEMO package [24] were randomly mixed linearly. The properties are likely to be known a priori. We propose to use
CICA-R algorithm was used to separate the speech signals. simple reference functions, assuming that the frequency of
The separated (normalized) signals by all the algorithms are the signals are known. Unlike earlier approaches to CICA-R,
shown in figure 2. Table 3 gives the SNR values of signals the present algorithm has the advantage of being suitable for
separated by different algorithms. The Performance Index both circular and non-circular sources.
values are 0.54(±0.20) for ICA, 0.67(±0.13) for Fast CICA,
0.59(±0.11) for CMN, and 0.48(±0.13) for CICA-R. The R EFERENCES
CICA-R showed the highest SNR values and the lowest PI
values, indicating its effectiveness separating speech signals [1] P. Comon, “Independent Component Analysis: a new concept?” Signal
in a cocktail party problem. Processing, Elsevier, vol. 36, pp. 287–314, 1994.
[2] A. Cichocki and S. Amari, Adaptive Blind Signal and Image Process-
ing. John Wiley, Sons, Inc., 2002.
V. D ISCUSSION [3] A. Hyvarinen, J. Karhunen, and E. Oja, Independent Component
We proposed CICA-R to extend the ICA-R framework Analysis. John Wiley, Sons, Inc., 2001.
[4] E. Bingham and A. Hyvarinen, “ A Fast Fixed-Point Algorithm
to handle complex signals. The CICA-R was shown to For Independent Component Analysis of Complex Valued Signals,”
more effective in extracting complex sources when a pri- International Journal of Neural Systems, vol. 10, pp. 1–8, 2000.
ori information of the desired signals are available in the [5] K. Raju, T. Ristaniemi, Karhunen, and E. Oja, “Suppression of
bitpulsed jammer signals in ds-cdma array system using independent
form of reference signals than Complex ICA (CICA). In component analysis,” in Proceedings of International Symposium on
experiments with synthetic data, CICA-R performs better Circuits and Systems, vol. 1, pp. 189–192, 2002.
than ICA, Fast CICA, or CMN. As CICA-R produces only [6] B. Karlsen, H. B. Sorensen, J. Larsen, and K. B. Jackobsen, “Inde-
pendent component analysis for clutter reduction in ground penetrating
the desired signals, it is computationally more efficient, radar data,” in Proceedings of SPIE AeroSense, vol. 4742, pp. 378–
389, 2002.
[7] J. Anemullera, T. J. Sejnowskia, and S. Makeig, “Complex indepen-
TABLE III dent component analysis of frequency-domain electroencephalographic
T HE SNR VALUES ( D B) OF THE SEPARATED SPEECH SIGNALS BY ICA, data,” Neural Networks, vol. 16, pp. 1311–1323, 2003.
FAST CICA, CMN, AND CICA-R METHODS IN THE COCKTAIL PARTY [8] J. Eriksson and V. Koivunen, “Complex-valued ICA using second
order statistics,” in Proceedings of IEEE Workshop Machine Learning
PROBLEM
Signal Processing, pp. 183–192, 2004.
source ICA Fast CICA CMN CICA-R [9] M. Novey and T. Adali, “Complex ICA by Negentropy Maximization,”
S1 8.98±0.86 1.26±1.04 2.41±1.99 24.42±3.17 IEEE Transactions Neural Networks, vol. 19, No. 4, pp. 596–609,
S2 3.04±1.12 15.43±2.58 21.06±3.20 37.89±2.96 2008.
S3 8.30±0.99 4.24±1.80 16.32±2.93 26.87±2.68 [10] J.-F. Cardoso and A. Souloumiac, “Blind beamforming for non-
S4 17.69±1.30 10.12±2.31 1.77±1.06 24.81±1.75 Gaussian signals,” Institute Electrical Engineering Proceedings-Radar
Signal Processing, vol. 140, pp. 362–370, 1993.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on January 21,2025 at 07:31:25 UTC from IEEE Xplore. Restrictions apply.
 (a) (b) (c)

(d) (e) (f)

Fig. 1. Experiments on synthetic data of convoluted mixtures: (a) original source signals; (b) source signals extracted by Fast CICA; (c) source signals
extracted by ICA-R; (d) mixed inputs; (e) source signals extracted by CICA-R; (f) source signals extracted by CMN algorithms.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on January 21,2025 at 07:31:25 UTC from IEEE Xplore. Restrictions apply.
 (a) (b)

(c) (d)
Fig. 2. The separation of speech signals in the cocktail party problem: (a) original source signals; (b) extracted signals by CICA-R; (c) extracted signals
by Fast CICA; and (d) extracted signals by CMN algorithm.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on January 21,2025 at 07:31:25 UTC from IEEE Xplore. Restrictions apply.
[11] W. Lu and J. C. Rajapakse, “ICA with reference,” in Proceedings of
Third International Conference on ICA and Blind Source Separation,
pp. 120–125, 2001.
[12] ——, “ICA with Reference,” Neurocomputing, vol. 69, pp. 2244–2257,
2006.
[13] ——, “Constrained Independent Component Analysis,” in Advance in
Neural Information Processing Systems, T. Leen, T. Dietterich, and
V. Tresp, Eds. MIT Press, 2003, vol. 13, pp. 570–576.
[14] ——, “Approach and Applications of Constrained ICA,” IEEE Trans-
actions on Neural Networks, vol. 16, pp. 203–212, 2005.
[15] Q. H. Lin, Y. R. Zheng, F. L. Yin, H. Liang, and V. D. Calhoun, “ A
Fast Algorithm for one-unit ICA-R,” Information Sciences, vol. 177,
pp. 1265–1275, 2007.
[16] L. K. L. Joshua and J. C. Rajapakse, “Extraction of event-related
potentials from EEG signals using ICA with reference,” in Proceedings
of International Joint Conference on Neural Networks, pp. 2526–2531,
2005.
[17] C. W. Hesse and C. J. James, “On semi-blind source separation
using spatial constraints with applications in EEG analysis,” IEEE
Transactions on Biomedical Engineering, vol. 53, pp. 2525–2534,
2006.
[18] J. Li and Q.-H. Lin, “One-Unit Fixed-Point Complex-valued ICA-
R Algorithm Using Magnitude Information(in Chinese),” Journal of
Electronics & Information Technology, vol. 30 No. 11, pp. 2666–2669,
2008.
[19] T. Azetsu, E. Uchino, and N. Suetake, “Blind separation and sound
localization by using frequency-domain ICA,” Soft Computing, vol. 11,
pp. 185–192, 2007.
[20] W. Lu and J. C. Rajapakse, “Eliminating Indeterminacy in ICA,”
Neurocomputing, vol. 50, pp. 271–290, 2003.
[21] R. Prasad, H. Saruwatari, and K. Shikano, “An ICA Algorithm for
Separation of Convolutive Mixture of Speech Signals,” International
Journal of Information Technology, vol. 2, No.4, pp. 273–283, 2004.
[22] S. Nawab and T. Quatieri, Short-Time Fourier Transform. Pearson
Education,Inc., 1987.
[23] H. Li and T. Adali, “A Class of Complex ICA Algorithms Based on
the Kurtosis Cost Function,” IEEE Transactions on Neural Networks,
vol. 19, pp. 408–420, 2008.
[24] R. Prasad, H. Saruwatari, A. Lee, and K. Shikano, “A fixed point
ICA algorithm for convoluted speech separation,” in Proceedings of
International Symposium on ICA and BSS, pp. 579–584, 2003.
[25] W. Xiong, Y.-O. Li, H. Li, T. Adali, and V. D. Calhoun, “On
ICA of Complex-valued fMRI: Advantages and Order Selection,” in
Proceedings of IEEE International Conference on Acoustics, Speech,
and Signal Processing, pp. 529–532, 2008.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on January 21,2025 at 07:31:25 UTC from IEEE Xplore. Restrictions apply.