独立成分解析
DESCRIPTION
九州工業大学が出版。TRANSCRIPT
-
16 5 25
1(Independent Component Analysis:ICA) 198090
ICA ICA ICA
1
ICA 1980 Jutten Herault 2
Comon[3], Amari & Cardoso[4] ICA Comon
(PCA) projection persuit projection persuit ICA ICA
ICA
1
2[1] [2] ICA
ICAEEG [5] fMRI [6]
[7, 8]
[9, 10]
J-F. Cardoso JADE [11] A. Hyvarinen FastICA [12] ICA central web page [13] Matlab ()
ICA
ICA
ICAICA
1
-
ICA
2
2.1 ICA
s(t) = (s1(t), . . . , sn(t))T
t = 0, 1, 2, . . .
s(t) 0T
x(t) = (x1(t), . . . , xm(t))T
t = 0, 1, 2, . . .
x m n
s(t) x(t)
x(t) = As(t), (1)
( 1)A m n ICA s(t) A x(t) n
s
s
x
x
1
2
1
2
11a
a21
a 22
a 12
1: 2 2 ICA
n m nm W
y(t) = Wx(t), (2)
y(t)WA = I(I n n ) y(t) s(t) y(t)
WA = PD (P 1 n n D n n )
2.2
ICA
3 [14, 15]
JADE FastICA ICA
si ()
W
x y
y
p(y) = p(y1, , yn)
W
y yi p(yi) yi
p(y) =n
i=1
p(yi)
p(y)n
i=1 p(yi) W
p(y) n
i=1 p(yi) Kullback-Leibler ()KL
KL(W ) =
p(y) logp(y)n
i=1 p(yi)dy (3)
= H(Y ;W ) +n
i=1
H(Yi;W ).
H(Y ;W ) H(Yi;W ) p(y)dy = p(x)dxp(y) = p(x)/|W | (|W | W )H(Y ;W ) H(Yi;W ) p(x) W
2
-
H(Y ;W ) =
p(y) logp(y)dy
=
p(x)(log p(x) log |W |)dx
= H(X) + log |W |,
H(Yi;W ) =
p(y) logp(yi)dy
=
p(x) log p(yi)dx
KL(W ) p(yi) 0 KL(W )
W KL(W )W
=((WT )1 Ex[(y)xT ]
)=(I Ex[(y)yT ]
)(WT )1 (4)
(y) = (
log p(y1)y1
, . . . , log p(yn)
yn
)T
W (4) (WT )1
WTW [16]
W (I Ex[(y)yT ])W (5)
[17]
W
Wt+1 = Wt + (I (y)yT)Wt.
(y)
W [4]W (y) ICA
x(t)
x(t)x(t + )T
= A
s(t)s(t+ )T
AT
= A
Rs1() 0. . .
0 Rsn()
AT ,
x(t) Rsi() si(t) s(t)
si(t) W y(t)
y(t)y(t+ )T
=(WAs(t)) (WAs(t+ ))T
=
21Rs1 () 0. . .
0 2nRsn ()
,
1, 2, . . . , n 1, 2, . . ., n i
W
y(t) x(t) i W
[18]
W x(t)x(t+ i)T WT = i, i = 1, . . . , r, (6)
i 2
W [18]2 i
[15]
KL [19]
3
-
si(t) [20]
Q(t) =12
{n
i=1
logE[y2i (t)] log detE[y(t)y(t)T ]}
,
(7)
(E[] ) Q ICA
Q E[y(t)y(t)T ] 0 W
W Q 0
Q
W
3
ICA
ICA
3.1
MEG [21, 22]MEG 100200MEG (1msec) ( mm)
(1014T)MEG () ( 1)
1
T Hz
1014 a few 20
1045 a few 10() 1011 1013 50 or 60
1014
1:
Low-Pass
50100
505
fT
x1
505
x2
100 0 100 200 300 4005
05
time(msec)
x3
2: MEG : 100 120 3 0 100msec 412msec
100 2 0msec
ICA x
x = As
1 s
4
-
ICA
x = As+
x = As+ = As = (A, I)
(s
)
ICA
[22] ICA A
ICA 3
[23, 24]
3.2
3
ICA MEG
(artifact) artifact
3 y9 180Hz
y11
412msec
y1
y2
y3
y4
y5
y6
y7
y8
y9
y10
y11
y12
y13
y14
y15
y16
100 0 100 200 300 400time(msec)
y17
3:
3
505
fT
x1
505
x2
100 0 100 200 300 4005
05
time(msec)
x3
4:
5
-
MEG 4 2 (100msec0msec) ( 80msec150msec)
MEG [25] 5 ICA 4
fMRI
5
()ICA
2 150msec 80150msec
V1
ICA
ICA
ICA
ICA
5: MEG
4
ICA
ICA
MEG, EEG, fMRI
6
-
x(t) = A(t) s(t)xi(t) =
k
aik(t) sk(t),
aik(t) sk(t) =
=0
aik()sk(t ),
( 6)
(FIR)
s
s
x
x
1
2
1
2
21
22
11
12
a (t)a
(t)
a (t)
a (t)
6: 2 2
W y(t) W (t)
y(t) = W (t) x(t)W (t) = {wij(t)}
FIR
(4) (7) [10, 26]
[27, 15, 28]x(t), s(t), A(t) Fourier
x(), s(), A()
x() = A()s(),
msec
x(, ts) = A()s(, ts),
x(, ts) s(, ts) x(t), s(t) windowed Fourier
(1) (1)
ICA
W () W ()1 A() W () W ()1
[28, 15]
ICA
5
ICA
34 ICAICA
7
-
ICA
ICA
s A
s
si
2
2 2
[29]
A
A [30, 8]
ICA
[31]KL [32, 33]
[34]
6
ICA web
[35, 36, 37]3 [38, 39]
ICA ICA 1999 1
1 3
ICAICA2001 [40]
2003 4 4 ICA [41]
7
[1] Jutten, C. and Herault, J. (1991): Separation of
sources, Part I, Signal Processing, Vol. 24, No. 1,pp. 110.
[2] Jutten, C. and Taleb, A.: Source separation:from dusk till dawn, in Proceedings of Interna-tional Workshop on Independent Component Analy-sis and Blind Signal Separation (ICA2000), pp. 1526 (2000).
[3] Comon, P. (1994): Independent component analysis,A new concept?, Signal Processing, Vol. 36, No. 3,pp. 287314.
[4] Amari, S. and Cardoso, J.-F. (1997): Blind SourceSeparation Semiparametric Statistical Approach,
3
8
-
IEEE Transactions on Signal Processing, Vol. 45,No. 11, pp. 26922700.
[5] Makeig, S., Jung, T.-P., Bell, A. J., Ghahremani, D.and Sejnowski, T. J. (1997): Blind Separation ofAuditory Event-related Brain Responses into Inde-pendent Components, Proc. Natl. Acad. Sci. USA,pp. 1097910984.
[6] Mckeown, M. J., Jung, T.-P., Makeig, S., Brown, G.,Kindermann, S. S., Lee, T.-W. and Sejnowski, T. J.:Spatially independent activity patterns in functionalmagnetic resonance imaging data during the Stroopcolor-naming task, in Proceedings of the NationalAcademy of Sciences, Vol. 95, pp. 803810 (1998).
[7] Bell, A. J. and Sejnowski, T. J. (1997): The inde-pendent component of natural scenes are edge l-ters, Vison Research, Vol. 37, pp. 33273338.
[8] Olshausen, B. A. and Field, D. J. (1996): Emergenceof simple-cell receptive eld properties by learninga sparce code for natural images, Nature, Vol. 381,pp. 607609.
[9] Bell, A. J. and Sejnowski, T. J. (1995): An informa-tion maximization approach to blind separation andblind deconvolution, Neural Computation, Vol. 7,No. 6, pp. 11291159.
[10] Douglas, S. C. and Cichocki, A. (1997): NeuralNetworks for Blind Decorrelation of Signals, IEEETransactions on Signal Processing, Vol. 45, No. 11,pp. 28292842.
[11] http://www.tsi.enst.fr/icacentral/Algos/cardoso/.
[12] http://www.cis.hut./projects/ica/fastica/.
[13] http://www.tsi.enst.fr/icacentral/.
[14] Cardoso, J.-F.: The three easy routes to indepen-dent component analysis; contrasts and geometry,in Proceedings of International Workshop on Inde-pendent Component Analysis and Blind Signal Sep-aration (ICA2001), pp. 16 (2001).
[15] Murata, N., Ikeda, S. and Ziehe, A. (2001): An Ap-proach to Blind Source Separation Based on Tem-poral Structure of Speech Signals, Neurocomputing,Vol. 41, No. 1-4, pp. 124.
[16] Amari, S., Cichocki, A. and Yang, H. H.: A NewLearning Algorightm for Blind Signal Separation, inTouretzky, D. S., Mozer, M. C. and Hasselmo, M. E.eds., Advances in Neural Information ProcessingSystems, Vol. 8, pp. 757763, The MIT Press, Cam-bridge MA (1996).
[17] Amari, S., Chen, T. and Cichocki, A. (1997): Stabil-ity Analysis of Learning Algorithms for Blind SourceSeparation, Neural Networks, Vol. 10, No. 8, pp.13451351.
[18] Molgedey, L. and Schuster, H. G. (1994): Separationof a mixture of independent signals using time de-layed correlations, Phys. Rev. Lett., Vol. 72, No. 23,pp. 36343637.
[19] Amari, S.: ICA of temporally correlated signals learning algorithm, in Proceedings of InternationalWorkshop on Independent Component Analysis andBlind Signal Separation (ICA99), pp. 1318 (1999).
[20] Matsuoka, K., Ohya, M. and Kawamoto, M. (1995):A Neural Net for Blind Separation of NonstationarySignals, Neural Networks, Vol. 8, No. 3, pp. 411419.
[21] Ikeda, S.: ICA on Noisy Data: A Factor AnalysisApproach, in Girolami, M. ed., Advances in Inde-pendent Component Analysis, chapter 11, pp. 201215, Springer-Verlag London Ltd. (2000).
[22] Ikeda, S. and Toyama, K. (2000): Independent Com-ponent Analysis for Noisy Data MEG data analy-sis, Neural Networks, Vol. 13, No. 10, pp. 10631074.
[23] Cao, J., Murata, N., Amari, S., Cichocki, A. andTakeda, T.: A Robust ICA Approach for unaver-aged single-trial auditory evoked elds data decom-position, in Proceedings of International Workshopon Independent Component Analysis and Blind Sig-nal Separation (ICA2001), pp. 445450 (2001).
[24] Kawanabe, M. and Murata, N.: Independent Com-ponent Analysis in the Presence of Gaussian noisebased on estimating functions, in Proceedings of In-ternational Workshop on Independent ComponentAnalysis and Blind Signal Separation (ICA2000),pp. 3944 (2000).
[25] Toyama, K., Yoshikawa, K., Yoshida, Y., Kondo, Y.,Tomita, S., Takanashi, Y., Ejima, Y. andYoshizawa, S. (1999): A new method for magnetoen-cephalography: A three dimensional magnetometer-spatial lter system, Neuroscience, Vol. 91, No. 2,pp. 405415.
[26] Kawamoto, M., Matsuoka, K. and Ohnishi, N.(1998): A method of blind separation for convolvednon-stationary signals, Neurocomputing, Vol. 22,No. 1-3, pp. 157171.
[27] Smaragdis, P. (1998): Blind separation of convolvedmixtures in the frequency domain, Neurocomputing,Vol. 22, No. 1-3, pp. 2134.
[28] Ikeda, S. and Murata, N.: A method of blind separa-tion based on temporal structure of signals, in Pro-ceedings of 1998 International Conference on Neu-ral Information Processing (ICONIP98), Vol. 2, pp.737742, Kitakyushu, Japan (1998).
[29] Rickard, S., Balan, R. and Rosca, J.: Real-TimeTime-Frequency Based Blind Source Separation, inProceedings of International Workshop on Indepen-dent Component Analysis and Blind Signal Separa-tion (ICA2001), pp. 651656 (2001).
[30] Hyvarinen, A., Hoyer, P. and Inki, M. (2001): To-pographic Independent Component Analysis, Neu-ral Computation, Vol. 13, No. 7, pp. 15251558.
9
-
[31] Bach, F. R. and Jordan, M. I.: Kernel In-dependent Component Analysis, Technical Re-port UCB//CSD-01-1166, University of California,Berkeley, (2001).
[32] Matsuyama, Y., Katsumata, N. and Imahara, S.:Convex Divergence as a Surrogate Function for Inde-pendence: The f-Divergence ICA, in Proceedings ofInternational Workshop on Independent ComponentAnalysis and Blind Signal Separation (ICA2001),pp. 3136 (2001).
[33] Minami, M. and Eguchi, S.: Robust Blind SourceSeparation by -Divergence, to appear in NeuralComputation (2002).
[34] Akuzawa, T.: New Fast Factorization Method forMultivariate Optimization and its Realization asICA Algorithm, in Proceedings of InternationalWorkshop on Independent Component Analysis andBlind Signal Separation (ICA2001), pp. 114119(2001).
[35] Lee, T.-W. (1998): Independent Component Analy-sis, Kluwer Academic Publishers.
[36] Hyvarinen, A., Karhunen, J. and Oja, E. (2001): In-dependent Component Analysis, John Wiley & Sons,Inc.
[37] Cichocki, A. and Amari, S. (2002): Adaptive BlindSignal and Image Processing, John Wiley & Sons,Inc.
[38] Girolami, M. ed. (2000): Advances in IndependentComponent Analysis, Springer-Verlag London Ltd.
[39] Haykin, S. ed. (2000): Unsupervised Adaptive Fil-tering, Volume 1, Blind Source Separation, Vol. 1,John Wiley & Sons, Inc.
[40] http://www.ica2001.org.
[41] http://ica2003.jp.
10