2013 11 01(fast_grbf-nmf)_for_share
TRANSCRIPT
Smooth Non-negative Matrix Factorization
using Function Approximation
November 1, 2013
Lab. Seminar
Tatsuya YokotaRIKEN BSI
1
2Our interest
BIG DATA
[from IBM analytics website]
Visualization
Feature analysis
ClusteringClassification
With more precision !! With more fast !!
[3] Daniel D. Lee and H. Sebastian Seung (1999). "Learning the parts of
objects by non-negative matrix factorization". Nature 401 (6755): 788–791
There are many nonnegative real-valued data such as image,
text, and physical/chemical spectra.
NMF is a technique to decompose a nonnegative observed
matrix Y into two nonnegative matrices A and X.
Optimization criterion
3NMF: nonnegative matrix factorization
Update A Update X
Convergence !!
4Blind Source Separation by NMF
vectorization
Observed signal 1 Observed signal 2
Estimated Sources
NMF
5Clustering by NMF
・・・
・・・
・・・
・・・
・・・
Dataset Centroids Weight parameters
Visualization by PCA
We can obtain centroids and
its clusters at the same time!!
Penalty termFused penalty [Chen et al., 2006]
Gibbs regularization [Zdunek and Cichocki, 2007, 2008]
Function approximation (FA) [Zdunek, 2012]
Approximate a matrix A by function approximation
Φ is a set of smooth basis functions
W is a weight parameter matrix
6Smooth NMF
Function approximation
7Smooth basis function Φ
Approximate as smooth feature vector
×
For X
Smooth NMF (GRBF-NMF)
8Algorithm for Smooth NMF
Fixed matrix
optimize optimize
[Zdunek, 2012] Zdunek, Rafał. "Approximation of feature vectors in nonnegative matrix factorization
with Gaussian radial basis functions." Neural Information Processing. Springer Berlin Heidelberg, 2012.
NNLS problem (quickly-solvable)
For W
QP problem (slowly)
Convergence !!
Improvement algorithm very fast (fastGRBF)
Extension to tensor decomposition
9Our research
NNLS problem (quickly-solvable)
QP problem (slowly)
SLOWImprove by fast HALS
based algorithm
CP model Tucker model
gr
ar
cr
br
GA
C
BT
New Criterion
HALS algorithm (fast)
10FastGRBF-NMF
Update Update
For r = 1, …, R
Convergence !!
11Extension to Tensor
GA
C
BT
2-direcions
smoothing
GU
C
BT
GU
C
VT
1-direcion
smoothing
All-directions
smoothing
GU
W
VT
T
Dataset: Yale faces dataset
We used (65×51 pixels)×(10 images)
Salt and pepper noise added
12Result: Denoising
PSNR( , )
Denoising score
NMF
・・・
10 images
65x51
= 3315
pixels
Noise level 10 dB (10 images are used)
Compare (a)PSNR and (b)Computational time
13Result: Denoising
(a) PSNR
(b) Time
New
New
NewNew
Algorithm Multiplicative HALS GRBF fastGRBF
Ave. of times[s] 0.499 0.759 720.0 2.23
About 300 times faster
Toy problem
Evaluation criteria
14Result: Blind Source Separation
Linear
Mixing
System
Gaussian noise
(a) 5 smooth sources (b) 20 noisy observations (4.4 dB)
+ =NMF
SIR( , )
Source estimation
SIR( , )
Denoising
Result :
Toy problem
15Result: Blind Source Separation
New New
Original Estimator
Denoising Score Estimation Score
Result :
Far-infrared
Spectra
16Result: Blind Source Separation
NewNew
Original Gibbs Reg. fastGRBF
Denoising Score Estimation Score
CBCL faces:
we used (19×19 pixels) × 100 images
Added Gaussian noise
Matrix factorization model
Tenor factorization model
17Result: CBCL faces with Tensor model
・・・
100 images
19x19
= 361
pixels
Ordinary NMF
fastGRBF-NMF
100 images
19 pixels
19 pixels
GA
C
B TOrdinary Nonneg. Tucker Dcomp. (NTD)
Ordinary Nonneg. CP Dcomp. (NCPD)
GU
C
V T
fastGRBF-NTD-2way-smoothing
fastGRBF-NCPD-2way-smoothing
Result :
Matrix model
18Result: CBCL faces with matrix model
Noisy data 16.2 dB
3x3 smooth filter 22.8 dB
multipl. NMF 21.9 dB
fastGRBF-NMF 23.4 dB
Proposed methods
Ordinary methods
19Result: CBCL faces with tensor model
Result :
Tensor modelNTD 20.7 dB
NCPD 23.2 dB
fastGRBF-NTD 23.2 dB
fastGRBF-NCPD 23.3 dB
Proposed methods
Ordinary methods
20Result: 3d-tensor data denoising
7.21 dB
NCPD
(19.8 dB)
NTD
(13.5 dB)
fastGRBF-NCPD
(26.8 dB)
fastGRBF-NTD-
(23.9 dB)
Gaussian
noise
GA
C
B T
GU
W
V TT
Ordinary model
fastGRBF-NTD/NCPD-3way-smoothing
Summary
We proposed new fast algorithm for GRBF-NMF
It succeeded about 300 faster computational time
We extended its model to tensor decomposition model
Their method performed good results
Future work for CIFA
Common smoothness and individual sparseness
21Future work
Y1 ≒ AC A1 X1
Y2 ≒ AC A2 X2
Common smooth factor Individual sparse factor