Académique Documents
Professionnel Documents
Culture Documents
Separation
a.k.a. How You Will Definitively Enjoy Differently a Cocktail Party
Matthieu Puigt
Foundation for Research and Technology Hellas
Institute of Computer Science
mpuigt@forth.ics.gr
http://www.ics.forth.gr/~mpuigt
M. Puigt
April/May 2011
(1)
M. Puigt
April/May 2011
(1)
M. Puigt
April/May 2011
(1)
April/May 2011
"blabla"
"blibla"
"blabla"
Separation
"blabli"
"blibli"
"blibli"
Sources sj (n)
M. Puigt
Observations xi (n)
Outputs yk (n)
April/May 2011
"blibli"
Separation
"blabli"
"blabla"
"blibli"
Sources sj (n)
M. Puigt
Observations xi (n)
Outputs yk (n)
April/May 2011
Separation
"blabli"
"blabla"
"blibli"
"blibli"
Sources sj (n)
M. Puigt
Observations xi (n)
Outputs yk (n)
April/May 2011
Most of the approaches process linear mixtures which are divided in three
categories:
1
Linear instantaneous (LI) mixtures: xi (n) = Nj=1 aij sj (n) (Purpose of
this lecture)
M. Puigt
April/May 2011
Most of the approaches process linear mixtures which are divided in three
categories:
1
Linear instantaneous (LI) mixtures: xi (n) = Nj=1 aij sj (n) (Purpose of
this lecture)
2
s1 (n)
s2 (n)
x1 (n)
M. Puigt
x2 (n)
April/May 2011
Most of the approaches process linear mixtures which are divided in three
categories:
1
Linear instantaneous (LI) mixtures: xi (n) = Nj=1 aij sj (n) (Purpose of
this lecture)
2
3
s1 (n)
s2 (n)
x1 (n)
M. Puigt
x2 (n)
April/May 2011
Most of the approaches process linear mixtures which are divided in three
categories:
1
Linear instantaneous (LI) mixtures: xi (n) = Nj=1 aij sj (n) (Purpose of
this lecture)
2
3
s1 (n)
s2 (n)
x1 (n)
M. Puigt
x2 (n)
April/May 2011
A kind of magic?
Here, the operator is a simple matrix whose coefficients are unknown.
a11 s1 + a12 s2 = 5
a21 s1 + a22 s2 = 1
In Signal Processing, we do not have the unique above system of
equation but a series of such systems (due to samples)
We thus use the intrinsic properties of source signals to achieve the
separation (assumptions)
M. Puigt
April/May 2011
A kind of magic?
Here, the operator is a simple matrix whose coefficients are unknown.
a11 s1 + a12 s2 = 0
a21 s1 + a22 s2 = .24
In Signal Processing, we do not have the unique above system of
equation but a series of such systems (due to samples)
We thus use the intrinsic properties of source signals to achieve the
separation (assumptions)
M. Puigt
April/May 2011
A kind of magic?
Here, the operator is a simple matrix whose coefficients are unknown.
a11 s1 + a12 s2 = 4
a21 s1 + a22 s2 = 2
In Signal Processing, we do not have the unique above system of
equation but a series of such systems (due to samples)
We thus use the intrinsic properties of source signals to achieve the
separation (assumptions)
M. Puigt
April/May 2011
A kind of magic?
Here, the operator is a simple matrix whose coefficients are unknown.
a11 s1 (n) + a12 s2 (n) = x1 (n)
a21 s1 (n) + a22 s2 (n) = x2 (n)
In Signal Processing, we do not have the unique above system of
equation but a series of such systems (due to samples)
We thus use the intrinsic properties of source signals to achieve the
separation (assumptions)
M. Puigt
April/May 2011
M. Puigt
April/May 2011
April/May 2011
A generic problem
Many applications: biomedical, audio processing and audio coding,
telecommunications, astrophysics, image classification, underwater
acoustics, finance, etc.
M. Puigt
April/May 2011
M. Puigt
April/May 2011
Part I
Sparse Component Analysis
3
Underdetermined case
Conclusion
M. Puigt
April/May 2011
10
M. Puigt
April/May 2011
11
M. Puigt
April/May 2011
11
x2 (n0 )
x1 (n0 ) ,
we obtain:
x2 (n0 )
x1 (n0 )
0)
= aa21 ss1 (n
=
11 1 (n0 )
iT
a21
a11
h
21
Instead of [a11 , a21 ]T , we thus can estimate 1, aa11
Let us see why!
M. Puigt
April/May 2011
11
Imagine now that, for each source, we have (at least) one sample
(single-source samples) for which only one source is active:
a11 s1 (n ) + a12 s2 (n ) = x1 (n )
a21 s1 (n ) + a22 s2 (n ) = x2 (n )
M. Puigt
(2)
April/May 2011
12
Imagine now that, for each source, we have (at least) one sample
(single-source samples) for which only one source is active:
a11 s1 (n0 )
= x1 (n0 )
a21 s1 (n0 )
= x2 (n0 )
M. Puigt
(2)
April/May 2011
12
Imagine now that, for each source, we have (at least) one sample
(single-source samples) for which only one source is active:
a12 s2 (n1 ) = x1 (n1 )
a22 s2 (n1 ) = x2 (n1 )
M. Puigt
(2)
April/May 2011
12
Imagine now that, for each source, we have (at least) one sample
(single-source samples) for which only one source is active:
a11 s1 (n ) + a12 s2 (n ) = x1 (n )
a21 s1 (n ) + a22 s2 (n ) = x2 (n )
Ratio
x2 (n)
x1 (n)
(2)
a12
a12
a11
M. Puigt
April/May 2011
12
M. Puigt
April/May 2011
13
2
3
M. Puigt
April/May 2011
13
2
3
M. Puigt
April/May 2011
13
2
3
M. Puigt
April/May 2011
13
x2 (n)
x1 (n)
a21
a11
M. Puigt
x2 (n)
x1 (n)
April/May 2011
14
Estimating a new column (average over the considered zone of the ratio
x2
x1 )
Problem
Finding such zones is hard with:
continuous speech (non-civilized talk where everyone speaks at the
same time)
short pieces of music
M. Puigt
April/May 2011
15
Estimating a new column (average over the considered zone of the ratio
x2
x1 )
Problem
Finding such zones is hard with:
continuous speech (non-civilized talk where everyone speaks at the
same time)
short pieces of music
M. Puigt
April/May 2011
15
Estimating a new column (average over the considered zone of the ratio
x2
x1 )
Problem
Finding such zones is hard with:
continuous speech (non-civilized talk where everyone speaks at the
same time)
short pieces of music
M. Puigt
April/May 2011
15
Estimating a new column (average over the considered zone of the ratio
x2
x1 )
Problem
Finding such zones is hard with:
continuous speech (non-civilized talk where everyone speaks at the
same time)
short pieces of music
M. Puigt
April/May 2011
15
Estimating a new column (average over the considered zone of the ratio
x2
x1 )
Problem
Finding such zones is hard with:
continuous speech (non-civilized talk where everyone speaks at the
same time)
short pieces of music
M. Puigt
April/May 2011
15
Estimating a new column (average over the considered zone of the ratio
x2
x1 )
Problem
Finding such zones is hard with:
continuous speech (non-civilized talk where everyone speaks at the
same time)
short pieces of music
M. Puigt
April/May 2011
15
Estimating a new column (average over the considered zone of the ratio
x2
x1 )
Problem
Finding such zones is hard with:
continuous speech (non-civilized talk where everyone speaks at the
same time)
short pieces of music
M. Puigt
April/May 2011
15
Estimating a new column (average over the considered zone of the ratio
x2
x1 )
Problem
Finding such zones is hard with:
continuous speech (non-civilized talk where everyone speaks at the
same time)
short pieces of music
M. Puigt
April/May 2011
15
X1 () = a11 S1 () + a12 S2 ()
Previous TEMPROM approach still applies on Frequency domain.
M. Puigt
April/May 2011
16
STFT
z}|{
x1 (n) = a11 s1 (n) + a12 s2 (n) X1 (n, ) = a11 S1 (n, ) + a12 S2 (n, )
M. Puigt
April/May 2011
17
M. Puigt
April/May 2011
18
M. Puigt
April/May 2011
18
Audio examples
Simulation
4 real English-spoken signals
Mixed with:
1
0.9
A=
0.92
0.93
0.92
0.9
1
0.9
0.9
1
0.9
0.92
0.93
0.92
0.9
1
Observation 2
Observation 3
Observation 4
Output 2
Output 3
Output 4
Source 3
Source 4
Source 2
A very short introduction to BSS
April/May 2011
19
Observation 2
April/May 2011
21
2
3
4
Source 1
M. Puigt
Source 2
Output 2
Source 3
Output 3
April/May 2011
22
j>i
s
i=1
jb1 , . . . , jbP = arg min q P
i=1 Rsub
s (i,i)
Source 1
Source 2
Output 2
Source 3
Output 3
April/May 2011
22
Conclusion
Conclusion
1
Many methods based on the same stages propose improved criteria for
finding single-source zones and estimating the mixing parameters
M. Puigt
April/May 2011
23
References
F. Abrard, Y. Deville, and P. White: From blind source separation to blind source cancellation in the
underdetermined case: a new approach based on time-frequency analysis, Proc. ICA 2001, pp.
734739, San Diego, California, Dec. 913, 2001
F. Abrard and Y. Deville: A time-frequency blind signal separation method applicable to
underdetermined mixtures of dependent sources, Signal Processing, 85(7):1389-1403, July 2005.
P. Bofill and M. Zibulevsky: Underdetermined Blind Source Separation using Sparse Representations,
Signal Processing, 81(11):23532362, 2001
Y. Deville, M. Puigt, and B. Albouy: Time-frequency blind signal separation: extended methods,
performance evaluation for speech sources, Proc. of IEEE IJCNN 2004, pp. 255-260, Budapest,
Hungary, 25-29 July 2004.
H. Mohimani, M. Babaie-Zadeh, and C. Jutten: A fast approach for overcomplete sparse
decomposition based on smoothed L0 norm, IEEE Transactions on Signal Processing, 57(1):289-301,
January 2009.
M. Puigt and Y. Deville: Iterative-Shift Cluster-Based Time-Frequency BSS for Fractional-Time-Delay
Mixtures, Proc. of ICA 2009, vol. LNCS 5441, pp. 306-313, Paraty, Brazil, March 15-18, 2009.
B. Sturm: Sparse approximation and atomic decomposition: considering atom interactions in
evaluating and building signal representations, Ph.D. dissertation, 2009.
http://www.mat.ucsb.edu/~b.sturm/PhD/Dissertation.pdf
E. Vincent: Complex nonconvex lp norm minimization for underdetermined source separation,Proc. of
ICA 2007, September 2007, London, United Kingdom. pp. 430-437
M. Xiao, S.L. Xie, and Y.L. Fu: A statistically sparse decomposition principle for underdetermined
blind source separation, Proc. ISPACS 2005, pp. 165168, 2005
O. Yilmaz and S. Rickard: Blind separation of speech mixtures via time-frequency masking, IEEE
Transactions on Signal Processing , 52(7):18301847, 2004.
M. Puigt
April/May 2011
24
Part II
Independent Component Analysis
This part is partly inspired by F. Theis online tutorials.
http://www.biologie.uni-regensburg.de/Biophysik/
Theis/teaching.html
7
10
11
Conclusion
M. Puigt
April/May 2011
25
RP fx (x)dx = 1
transformation: fAx (x) = | det(A)|1 fx (A1 x)
M. Puigt
April/May 2011
26
kurtosis
the combination of moments kurt(x) , E{x4 } 3(E{x2 })2 is called
kurtosis of x
kurt(x) = 0 if x Gaussian, < 0 if sub-Gaussian and > 0 if super-Gaussian
(speech is usually modeled by a Laplacian distribution = super-Gaussian)
sampling
in practice density is unknown only some samples i.e. values of random
function are given
given independent (xi )i=1,...,P with same density f , then x1 (), . . . , xn ()
for some event are called i.i.d. samples of f
strong theorem of large numbers: given a pairwise i.i.d. sequence (xi )iN
in L1 (), then (for almost all )
!
1 P
lim
xi () E{x1 } = 0
P+ P
i=1
M. Puigt
April/May 2011
27
negentropy
negentropy of x is defined by J(x) , H(xgauss ) H(x)
transformation: J(Ax) = J(x)
1
1
E{x3 }2 + 48
kurt(x)2
approximation: J(x) ' 12
information
I (x) , Pi=1 (H(xi )) H(x) is called mutual information of X
I (x) 0 and I (x) = 0 if and only if x is independent
transformation: I (x + c) = I (x) for scaling , permutation , and
translation c RP
M. Puigt
April/May 2011
28
April/May 2011
29
April/May 2011
29
Algebraic algorithm
eigenvalue decomposition
calculate eigenvectors and eigenvalues of C , Cov(x) i.e. search for
v RP \{0} with Cv = v
there exists an orthonormal basis {v1 , . . . , vP } of eigenvectors of C with
corresponding eigenvalues 1 , . . . , P
1 0
0
V T CV
=D
April/May 2011
30
M. Puigt
April/May 2011
31
PCA does half the job and we need to rotate the data to achieve the
separation!
M. Puigt
April/May 2011
31
ai
i
(i si )
orthogonal A
by assumption Cov(s) = Cov(x) = I
hence I = Cov(x) = ACov(s)AT = AAT
M. Puigt
April/May 2011
32
ICA algorithms
basic scheme of ICA algorithms (case P = N)
search for invertible W Gl(N) that minimizes some dependence
measure of WX
For example minimize mutual information I (Wx) (Comon, 1994)
Or maximize neural network output entropy H(f (Wx)) (Bell and
Sejnowski, 1995)
Earliest algorithm: extend PCA by performing nonlinear decorrelation
(Hrault and Jutten, 1986)
Geometric approach, seeing the mixture distributions as a parallelogram
whose directions are given by the mixing matrix columns (Theis et al.,
2003)
Etc...
M. Puigt
April/May 2011
33
Why?
Theorem of central limit states that sum of random variables tends to a
Gaussian distribution
M. Puigt
April/May 2011
34
Why?
Theorem of central limit states that sum of random variables tends to a
Gaussian distribution
Demixing systems Non-Gaussian output signals (at most one
Gaussian source (Comon, 1994))
Non-Gaussianity measures:
Kurtosis (kurt(x) = 0 if x Gaussian, > 0 if X Laplacian (speech))
Neguentropy (always 0 and = 0 when Gaussian)
Basic idea: given x = As, construct ICA matrix W, which ideally equals
A1
Recover only one source: search for b RN with y = bT x = bT As , qT s
Ideally b is row of A1 , so q = ei
Thanks to central limit theorem y = qT s is more Gaussian than all source
components si
At ICA solutions y ' si , hence solutions are least Gaussian
April/May 2011
34
Source distributions
M. Puigt
April/May 2011
35
M. Puigt
April/May 2011
35
M. Puigt
April/May 2011
36
FastICA Algorithm
s is not known after whitening underlinez = Vx search for w RN
with wT z maximal non-gaussian
because of q = (VA)T w we get |q|2 = qT q = (wT VA)(AT V T w) = |w|2
so if q S N1 also w S N1
(kurtosis maximization): Maximize w 7 |kurt(wT z)| on S N1 after
whitening.
M. Puigt
April/May 2011
37
Maximization
Algorithmic maximization by gradient ascent:
A differentiable function f : RN R can be maximized by local updates
in directions of its gradient
Sufficiently small learning rate > 0 and a starting point x(0) RN , local
maxima of f can be found by iterating x(t + 1) = x(t) + x(t) with
f
(x(t)) the gradient of f at x(t)
x(t) = grad f (x(t)) = x
in our case
grad|kurt(wT z)|(w) = 4sgn(kurt(wT z))(E(z{wT z)3 } 3|w|2 w)
Algorithm (gradient ascent kurtosis maximization):
Choose > 0 and w(0) S N1 .
Then iterate
w(t) , sgn(kurt(w(t)T z))E{z(w(t)T z)3 }
v(t + 1) , w(t) + w(t)
w(t + 1) ,
M. Puigt
v(t + 1)
|v(t + 1)|
April/May 2011
38
v(t + 1)
|v(t + 1)|
advantages:
higher convergence speed (cubic instead of quadratic)
parameter-free algorithm (apart from the starting vector)
FastICA (Hyvrinen and Oja, 1997)
M. Puigt
April/May 2011
39
uq (t + 1) , vq (t + 1)
uq (t+1)
|uq (t+1)|
Normalize wq (t + 1) ,
April/May 2011
40
M. Puigt
April/May 2011
41
M. Puigt
April/May 2011
41
M. Puigt
April/May 2011
41
M. Puigt
April/May 2011
42
AMUSE
bilinearity of autocovariance:
Rx () = E{x(t + )x(t)T } =
ARs (0)AT + 2 I
ARs ()AT
if = 0
if 6= 0
So symmetrized autocovariance Rx () , 12 Rx () + Rx ()T fulfills (for
6= 0)
Rx () = ARs ()AT
identifiability:
A can only be found up to permutation and scaling (classical BSS
indeterminacies)
if there exists Rs () with pairwise different eigenvalues no more
indeterminacies
M. Puigt
April/May 2011
43
April/May 2011
44
It is not so simple...
Dependent
Dependent (high correlation)
M. Puigt
Independent
???
Signal length
Music signals
(Abrard & Deville, 2003)
April/May 2011
45
It is not so simple...
???
Dependent
Dependent (high correlation)
M. Puigt
Independent
???
Signal length
Music signals
(Abrard & Deville, 2003)
April/May 2011
45
GI {s} =
bs (q)
det diag R
1 Q 1
log
bs (q)
Q q=1 2
det R
M. Puigt
April/May 2011
46
0.5
0.5
-2
10
-1
10
10
Excerpt duration (s)
10
-2
10
0.5
0.5
10-2
10-1
100
Excerpt duration (s)
101
0
10-2
M. Puigt
10-1
100
Excerpt duration (s)
10
10-2
10-1
100
Excerpt duration (s)
101
0.1
-0.1
-1
10
10
Excerpt duration (s)
101
-0.1
10-2
10-1
100
Excerpt duration (s)
101
April/May 2011
47
0.5
0.5
-2
10
-1
10
10
Excerpt duration (s)
10
-2
10
0.5
0.5
10-2
10-1
100
Excerpt duration (s)
101
0
10-2
10-1
100
Excerpt duration (s)
10
10-2
10-1
100
Excerpt duration (s)
101
0.1
-0.1
-1
10
10
Excerpt duration (s)
101
-0.1
10-2
10-1
100
Excerpt duration (s)
101
Speech signals
Dependent
Dependent
0
(more dependent than speech)
M. Puigt
Independent
Independent
Signal length
Music signals
April/May 2011
47
ICA Performance
Influence of dependency measures on the performance of ICA?
60 above pairs (30 speech + 30 music) of audio signals
Mixed by the Identity matrix
Demixed with Parallel FastICA & the Pham-Cardoso algorithm
80
80
60
60
40
40
20
20
10-2
10-1
100
Excerpt duration (s)
101
80
60
60
40
40
20
0
10-2
10-1
100
Excerpt duration (s)
101
20
10
-2
M. Puigt
-1
10
10
Excerpt duration (s)
10
10-2
10-1
100
Excerpt duration (s)
101
April/May 2011
48
ICA Performance
Influence of dependency measures on the performance of ICA?
60 above pairs (30 speech + 30 music) of audio signals
Mixed by the Identity matrix
Demixed with Parallel FastICA & the Pham-Cardoso algorithm
toto
toto
To conclude: behaviour linked to the size of time excerpt
1
Low size: music signals exhibit more dependencies than speech ones
M. Puigt
April/May 2011
48
Conclusion
Introduction to ICA
Historical and powerful class of methods for solving BSS
Independence assumption satisfied in many problems (including audio
signals if frames long enough)
Many criteria have been proposed and some of them are more powerfull
than others
ICA extended to Independent Subspace Analysis (Cardoso, 1998)
ICA can use extra-information about the sources Sparse ICA or
Non-negative ICA
Some available softwares:
FastICA: http://research.ics.tkk.fi/ica/fastica/
ICALab: http://www.bsp.brain.riken.go.jp/ICALAB/
ICA Central algorithms:
http://www.tsi.enst.fr/icacentral/algos.html
many others on personal webpages...
M. Puigt
April/May 2011
49
References
F. Abrard, and Y. Deville: Blind separation of dependent sources using the TIme-Frequency Ratio Of Mixtures approach, Proc. ISSPA 2003,
pp. 8184.
A. Bell and T. Sejnowski: An information-maximisation approach to blind separation and blind deconvolution, Neural Computation,
7:11291159, 1995.
A. Belouchrani, K. A. Meraim, J.F. Cardoso, and E. Moulines: A blind source separation technique based on second order statistics, IEEE
Trans. on Signal Processing, 45(2):434444, 1997.
J.F. Cardoso: Blind signal separation: statistical principles, Proc. of the IEEE, 9(10):20092025, Oct. 1998
P. Comon: Independent component analysis, a new concept?, Signal Processing, 36(3):287-314, Apr. 1994
J. Hrault and C. Jutten: Space or time adaptive signal processing by neural network models, Neural Networks for Computing, Proceedings of
the AIP Conference, pages 206211, New York, 1986. American Institute of Physics.
A. Hyvrinen, J. Karhunen, and E. Oja: Fast and Robust Fixed-Point Algorithms for Independent Component Analysis,IEEE Trans. on Neural
Networks 10(3)626634, 1999.
A. Kraskov, H. Stgbauer, and P. Grassberger: Estimating mutual information, Physical Review E, 69(6), preprint 066138, 2004.
L. Parra and P. Sajda: Blind Source Separation via generalized eigenvalue decomposition, Journal of Machine Learning Research,
(4):12611269, 2003.
D.T. Pham and J.F. Cardoso: Blind separation of instantaneous mixtures of nonstationary sources, IEEE Trans. on Signal Processing,
49(9)18371848, 2001.
M. Puigt, E. Vincent, and Y. Deville: Validity of the Independence Assumption for the Separation of Instantaneous and Convolutive Mixtures of
Speech and Music Sources, Proc. ICA 2009, vol. LNCS 5441, pp. 613-620, March 2009.
D. Smith, J. Lukasiak, and I.S. Burnett: An analysis of the limitations of blind signal separation application with speech, Signal Processing,
86(2):353359, 2006.
A. Souloumiac: Blind source detection and separation using second-order non-stationarity, Proc. ICASSP 1995, (3):19121915, May 1995.
F. Theis, A. Jung, C. Puntonet, and E. Lang: Linear geometric ICA: Fundamentals and algorithms, Neural Computation, 15:419439, 2003.
L. Tong, R.W. Liu, V. Soon, and Y.F. Huang: Indeterminacy and identifiability of blind identification, IEEE Transactions on Circuits and
Systems, 38:499509, 1991.
M. Puigt
April/May 2011
50
Part III
Non-negative Matrix Factorization
A really really short introduction to NMF
12
Whats that?
13
14
Conclusion
M. Puigt
April/May 2011
51
Whats that?
In many problems, non-negative observations (images, spectra, stocks,
etc) both the sources and the mixing matrix are positive
Goal of NMF: Decompose a non-negative matrix X as the product of
two non-negative matrices A and S with X = A S
Earliest methods developed in the mid90 in Finland, under the name of
Positive Matrix Factorization (Paatero and Tapper, 1994).
Famous method by Lee and Seung (1999, 2000) popularized the topic
Iterative approach that minimizes the divergence between X and A S
Xij
div(X|AS) = Xij log
Xij + (AS)ij .
(AS)ij
i,j
Non-unique solution, but uniqueness guaranteed if sparse sources (see e.g.
Hoyer, 2004, or Schachtner et al., 2009)
M. Puigt
April/May 2011
52
Why is it so popular?
Face recognition: NMF "recognizes" some natural parts of faces
S contains parts-based representation of the data and A is a weight matrix
(Lee and Seung, 1999).
M. Puigt
April/May 2011
53
M. Puigt
April/May 2011
54
M. Puigt
April/May 2011
55
Conclusion
Really really short introduction to NMF
For audio signals, basically consists in working in the Frequency
domain
Observations decomposed as a linear combination of basic waveforms
Need to cluster them then
M. Puigt
April/May 2011
56
References
P. Hoyer: Non-negative Matrix Factorization with Sparseness Constraints, Journal of Machine
Learning Research 5, pp. 14571469, 2004.
P. Paatero and U. Tapper: Positive matrix factorization: A non-negative factor model with optimal
utilization of error estimates of data values, Environmetrics 5:111126, 1994.
D.D. Lee and H.S. Seung:Learning the parts of objects by non-negative matrix factorization, Nature
401 (6755):788791, 1999.
D.D. Lee and H.S. Seung:Algorithms for Non-negative Matrix Factorization. Advances in Neural
Information Processing Systems 13: Proc. of the 2000 Conference. MIT Press. pp. 556562, 2001.
M.D. Plumbley: Conditions for non-negative independent component analysis, IEEE Signal
Processing Letters, 9(6):177180, 2002
R. Schachtner, G. Pppel, A. Tom, and E. Lang: Minimum Determinant Constraint for Non-negative
Matrix Factorization, Proc. of ICA 2009, LNCS 5441, pp. 106113, 2009.
T. Virtanen: Monaural Sound Source Separation by Non-Negative Matrix Factorization with Temporal
Continuity and Sparseness Criteria, IEEE Trans. on Audio, Speech, and Language Processing, vol 15,
no. 3, March 2007.
B. Wang and M.D. Plumbley: Musical audio stream separation by non-negative matrix factorization,
Proc. of the DMRN Summer Conference, Glasgow, 2324 July 2005.
M. Puigt
April/May 2011
57
Part IV
From cocktail party to the origin of
galaxies
From the paper:
M. Puigt, O. Bern, R. Guidara, Y. Deville, S. Hosseini, C. Joblin:
Cross-validation of blindly separated interstellar dust spectra, Proc. of
ECMS 2009, pp. 4148, Mondragon, Spain, July 8-10, 2009.
15
Problem Statement
16
17
Conclusion
M. Puigt
April/May 2011
58
M. Puigt
April/May 2011
59
M. Puigt
April/May 2011
60
Interstellar dust
Absorbs UV light and re-emit it in
the IR domain
Several grains in PhotoDissociation Regions (PDRs)
Spitzer IR spectrograph provides
hyperspectral datacubes
N
x(n,m) () = a(n,m),j sj ()
j=1
Polycyclic Aromatic
Hydrocarbons
Very Small Grains
Big grains
M. Puigt
April/May 2011
60
Interstellar dust
Absorbs UV light and re-emit it in
the IR domain
Several grains in PhotoDissociation Regions (PDRs)
Spitzer IR spectrograph provides
hyperspectral datacubes
N
x(n,m) () = a(n,m),j sj ()
j=1
April/May 2011
60
Separation
April/May 2011
61
FastICA:
Maximization of
non-Gaussianity
Sources are stationary
M. Puigt
April/May 2011
62
M. Puigt
62
M. Puigt
62
M. Puigt
62
Xij
b
b
b
b
div X|AS = Xij log
Xij + AS
ij
bb
i,j
A
S
ij
M. Puigt
April/May 2011
62
Pre-processing stage
Additive noise not taken into account in the mixing model
More observations than sources
Pre-processing stage for reducing the noise & the complexity:
M. Puigt
April/May 2011
63
c R. Croman www.rc-astro.com
M. Puigt
April/May 2011
64
c R. Croman www.rc-astro.com
M. Puigt
April/May 2011
64
c R. Croman www.rc-astro.com
M. Puigt
April/May 2011
64
c R. Croman www.rc-astro.com
April/May 2011
64
yk () = j sj ()
M. Puigt
April/May 2011
65
M. Puigt
April/May 2011
65
Conclusion
Conclusion
1
M. Puigt
April/May 2011
66