TENSORIZATION
AND APPLICATIONS IN
BLIND SIGNAL SEPARATION
OTTO DEBALS
Supervisors:
Prof. Lieven De Lathauwer
Prof. Marc Van Barel
Data
Patterns/
relations
Predictions
Self-driving car algorithms
Predictions:
behaviors
Traffic & congestion models
Predictions:
time to goal
Electrocardiogram analysis
Predictions:
disorders
Data
Patterns/
relations
Predictions
Data representation?
Mining tools?
VECTOR = ONEWAY ARRAY
22 24 29 23 25 25 26
5.1 2.7 9.4 8.8 1.7 2.2 3.4
[
]
Mon Tue Wed Thu Fri Sat Sun
R0 E40 R1 E19 E17 E314 R4
Temperature in °C
MATRIX = TWOWAY ARRAY
Mon Tue Wed Thu Fri Sat Sun
MATRIX = TWOWAY ARRAY
Mon Tue Wed Thu Fri Sat Sun
MATRIX = TWOWAY ARRAY
Mon Tue Wed Thu Fri Sat Sun
Leuven
22 24 29 23 25 25 26
21
23 22 21
23 23 22
27 29 28 26 26 27 26
30 30 31
32 32 31
32
London Paris AthensDay
Week
Day
TENSOR = MULTIWAY ARRAY
Sheet 1
Sheet 2
Sheet 3
IN TERMS OF
DATA MINING TOOLS
Matrix tools
Information
Information
Tensor tools
VERY POWERFUL!
WHAT EXACTLY IS
Data
Application of
tensor tools
Data
Application of
tensor tools
Naturally
Data
Application of
tensor tools
Naturally
Physical quantities in function of height x width x depth
Video data: height x width x time
Color image data, hyperspectral data, …
Data
Application of
tensor tools
Naturally
By experiment design
Recommendation systems: user x movie → user x movie
x month x topic x …
Biomedical signal processing: electrode x time → electrode x time
x subject x …
Face databases: height x width → height x width
x person x emoticon x …
Data
Application of
tensor tools
Naturally
By experiment design
&
Data
Application of
tensor tools
Tensorization
Naturally By experiment design & TensorizedData
Application of
tensor tools
Tensorization
Naturally By experiment design & Tensorized&
Without using additional data
In a meaningful way
MORE ADVANCED
EXAMPLE: HANKELIZATION
a b c d e
a
b
c
b
c
d
c
d
e
Hankel matrixMORE ADVANCED
EXAMPLE: HANKELIZATION
a b c d e
a
b
c
b
c
d
c
d
e
Hankel matrixMORE ADVANCED
EXAMPLE: HANKELIZATION
a b c d e
a
b
c
b
c
d
c
d
e
Hankel matrixVARIOUS TENSORIZATION
TECHNIQUES EXIST
Tensorization
overview
Single
vector
vectors
Sets of
Hankelization Segmentation Löwnerization Monomial relations Time-frequency & time-scale Moments & cumulants Adjacency tensors Score functions Hessian & Jacobian matrices Covariance matrices Piecewise outer product
And many
more
BLIND SIGNAL
Source signals
Mixing level
Observed signals
Source signals
Mixing level
Observed signals
Source signals
Mixing level
Observed signals
Assumption of
independence
allows the
recovery of
underlying components
Emission spectra of individual analytes
Observed emission spectra of the compound at different
excitation levels Influence of
excitation and concentrations
Emission spectra of individual analytes
Observed emission spectra of the compound at different
excitation levels Influence of
excitation and concentrations
Emission spectra of individual analytes
Observed emission spectra of the compound at different
excitation levels Influence of
excitation and concentrations
Assumption of
rational functions
allows
the recovery of
underlying components
Maternal and fetal
ECG signals Mixing level Observed ECG signals
Maternal and fetal
ECG signals Mixing level Observed ECG signals
Maternal and fetal
ECG signals Mixing level Observed ECG signals
Maternal and fetal
ECG signals Mixing level Observed ECG signals
Maternal and fetal
ECG signals Mixing level Observed ECG signals
Assumption of
independence / rational functions /
sums-of-Kronecker-products
allows the recovery of
underlying components
Maternal and fetal
ECG signals Mixing level Observed ECG signals
Assumption of
independence / rational functions /
sums-of-Kronecker-products
allows the recovery of
underlying components
HOW CAN WE USE
TENSORS
TO SOLVE THE PROBLEM OF
Translate assumptions
TENSORIZATION
TWO THINGS ARE NEEDED.
1
Different assumptions require different tensorization techniques: Independence
Rational functions Etc.
STATISTICS
LÖWNERIZATION Exponential polynomials HANKELIZATION
Translate assumptions
Identify underlying
components
TENSORIZATION
TENSOR TOOLS
TWO THINGS ARE NEEDED.
1
2
Different assumptions require different tensorization techniques: Independence
Rational functions Etc.
STATISTICS
LÖWNERIZATION Exponential polynomials HANKELIZATION
R0 8h
80
40
120
20
12h 16h 20h Measure of traffic Ring around BrusselsR0 E40 8h
80
40
120
20
20
10
30
5
12h 16h 20h Major Belgian highwayR0 E40 R1 8h
80
40
120
20
20
10
30
5
40
20
60
10
12h 16h 20h Ring around AntwerpR0 E40 R1 ChurchStr. 8h
80
40
120
20
20
10
30
5
40
20
60
10
4
2
6
1
12h 16h 20hRANK-1 MATRIX
R0 E40 R1 ChurchStr. 8h
80
40
120
20
20
10
30
5
40
20
60
10
4
2
6
1
12h 16h 20hRANK-1 MATRIX
R0 E40 R1 ChurchStr. 8h
80
40
120
20
20
10
30
5
40
20
60
10
4
2
6
1
12h 16h 20hx20
RANK-1 MATRIX
R0 E40 R1 ChurchStr. 8h
80
40
120
20
20
10
30
5
40
20
60
10
4
2
6
1
12h 16h 20h=
8h4
2
6
1
12h 16h 20h20 5
10
1
R0 E40 R1 ChurchStr.RANK-1 MATRIX
R0 E40 R1 ChurchStr. 8h
80
40
120
20
20
10
30
5
40
20
60
10
4
2
6
1
12h 16h 20h=
8h4
2
6
1
12h 16h 20h20
5
10
1
R0 E40 R1 ChurchStr.RANK-1 MATRIX
60 = 6 x 10
R0 E40 R1 ChurchStr. 8h
80
40
120
20
20
10
30
5
40
20
60
10
4
2
6
1
12h 16h 20h=
8h4
2
6
1
12h 16h 20h20 5
10
1
R0 E40 R1 ChurchStr.=
Normal behavior
=
Influence of major incident Normal behavior+
…
+
=
Influence of major incident Normal behavior+
…
+
Underlying components=
Influence of major incident Normal behavior+
…
+
=
Influence of major incident Normal behavior+
…
+
=
Normal behavior=
Influence of major incident Normal behavior+
…
+
=
+
…
+
Normal=
+
…
+
=
+
…
+
Normal
behavior major incidentInfluence of
Normal behavior and
=
+
…
+
=
+
…
+
Normal
behavior major incidentInfluence of Normal behavior and
TENSOR DECOMPOSITIONS ALLOW
THE
UNIQUE RECOVERY
OF
TENSOR-BASED BLIND SIGNAL SEPARATION
TENSOR-BASED BLIND SIGNAL SEPARATION
+ =
1
TENSORIZATIONTENSOR-BASED BLIND SIGNAL SEPARATION
+ =
1
TENSORIZATIONCOCKTAIL PARTY PROBLEM
TENSORIZATION
ONLY USEFUL
FOR SIGNAL SEPARATION?
AS A PARADIGM, FOR MUCH MORE!
Graph or data clustering Training neural networks Function approximations
Decoupling of systems Filter banks