robust kernel-based regression using orthogonal matching pursuit
DESCRIPTION
Presentetion for the IEEE International Workshop on MACHINE LEARNING FOR SIGNAL PROCESSING MLSP 2013 conference.TRANSCRIPT
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
ROBUST KERNEL-BASED REGRESSIONUSING ORTHOGONAL MATCHING PURSUIT
(OMP)
G. Papageorgiou, P. Bouboulis, S. Theodoridis
Department of Informatics and Telecommunications, National and KapodistrianUniversity of Athens, Greece
MLSP SEP. 22-25, 2013,Southampton, UK
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Outline
1 Introduction
2 Problem formulation
3 Unravelling KROMP
4 Experimental Results
5 Conclusions and future work
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Introduction
Let (yk , xk), k = 1, 2, ..., n, with yk ∈ R, xk ∈ Rm. A typicalregression task is to estimate the input-output relation satisfying:
yk = f (xk) + ηk , k = 1, 2, ..., n.
The problem is solved:
f linear: Least Squares in the Euclidean space.
f nonlinear: Assume that f ∈ H, where H will be assumed tobe a RKHS and the minimizer of
minf
n∑k=1
(yk − f (xk)
)2+ µ||f ||2H, µ > 0
admits a representation f̂ (x) =∑n
i=1 αiκ(x, xi ) (LS in H).
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Problem formulation
What if the noise samples contain outliers?
Is it possible to remove other type of noise too?
Let uk be impulse noise samples of unknown “energy”, assumed tobe outliers, hence sparse. Since LS is not robust against outlierswe reformulate:
Problem model
yk = f (xk) + uk + ηk , k = 1, 2, ..., n,
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Problem formulation
What if the noise samples contain outliers?
Is it possible to remove other type of noise too?
Let uk be impulse noise samples of unknown “energy”, assumed tobe outliers, hence sparse. Since LS is not robust against outlierswe reformulate:
Problem model
yk = f (xk) + uk + ηk , k = 1, 2, ..., n,
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Problem formulation
What if the noise samples contain outliers?
Is it possible to remove other type of noise too?
Let uk be impulse noise samples of unknown “energy”, assumed tobe outliers, hence sparse. Since LS is not robust against outlierswe reformulate:
Problem model
yk = f (xk) + uk + ηk , k = 1, 2, ..., n,
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Problem formulation
Assuming that the unknown function is expressed as a linearcombination of kernel functions of the formf =
∑nk=1 αkκ(·, xk) + c , we intend to solve:
Minimization Task
minα,c,u
‖u‖0s.t. ‖Kα + c1 + u− y‖22 + λ‖α‖22 + λc2 ≤ ε
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Problem formulation
Minimization Task
minα,c,u
||u||0 s.t. ||Kα + c1 + u− y||22 + λ||α||22 + λc2 ≤ ε
Remarks:
Keeps the square error low.
Regularization guards against overfitting.
Use of nonconvex function.
Exploits greedy sparse approximation algorithms from thefamily of (GA).
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Sparse approximation (denoising) algorithms
l0 (pseudo)-norm minimization:
minx||x||0 s.t. ||Ax− b||2 ≤ ε
Greedy Algorithms (GA)
(MP) (OMP) (OLS) (LARS)
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Sparse approximation (denoising) algorithms
l1 norm minimization:
minx||x||1 s.t. ||Ax− b||2 ≤ ε
or minx{λ||x||1 + 1
2 ||Ax− b||22} (
for appropriate Lagrangemultiplier λ = λ(A,b, ε)
)Homotopy(TNIPM)GradientProjection (GP)(IRLS)
IterativeShrinkage-ThresholdingProximalGradient (PG)
(CoSaMP)
AugmentedLagrangianMethods
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Column selection from matrix A
We recast the optimization task in matrix notation:
minz||u||0 s.t. ||Az− y||22 + λ||Bz||22 ≤ ε,
where A =[K 1 In
], z =
αcu
, B =
In 0 On
0T 1 0T
On 0 On
, meaning
that only the last part of z is sparse.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Column selection from matrix A
Initially (k := 0) we set A(0) = [K 1] and B(0) = In+1.
At each step:
We find one position (index) for an outlier (say 1 ≤ jk ≤ n).Then, A is augmented by the column jk of In and B isaugmented by zeros in order to match the appropriatedimensions (actually this is a projection matrix B2 = B).Finally, we solve minz ‖A(k)z− y‖22 + λ‖B(k)z‖22.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Column selection from matrix A
Green columns: columns from matrix A, which have not beenselected yet.Red columns: columns that have already been selected at previoussteps. These participate in the regularized minimization problem.
Step k := 0 : A(0) = [K 1]
Step k := 1 : A(1) = [K 1 e2]
Step k := 2 : A(2) = [K 1 e2 e3]
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Column selection from matrix A
Green columns: columns from matrix A, which have not beenselected yet.Red columns: columns that have already been selected at previoussteps. These participate in the regularized minimization problem.
Step k := 0 : A(0) = [K 1]
Step k := 1 : A(1) = [K 1 e2]
Step k := 2 : A(2) = [K 1 e2 e3]
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Column selection from matrix A
Green columns: columns from matrix A, which have not beenselected yet.Red columns: columns that have already been selected at previoussteps. These participate in the regularized minimization problem.
Step k := 0 : A(0) = [K 1]
Step k := 1 : A(1) = [K 1 e2]
Step k := 2 : A(2) = [K 1 e2 e3]
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Column selection from matrix A
Green columns: columns from matrix A, which have not beenselected yet.Red columns: columns that have already been selected at previoussteps. These participate in the regularized minimization problem.
Step k := 0 : A(0) = [K 1]
Step k := 1 : A(1) = [K 1 e2]
Step k := 2 : A(2) = [K 1 e2 e3]
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Kernel Regularized OMP (KROMP): Initialization
Select the kernel for the kernel matrix K , the regularizationparameter λ and the error threshold ε.
Step k := 0 : Solve the regularized LS problem and computethe residual z(0) := arg minz ||A(0)z− y||22 + λ||B(0)z||22,
r(0) = y − A(0)z(0), where S(0)act = {1, ..., n + 1},
A(0) = [K 1], B(0) = In+1.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Kernel Regularized OMP (KROMP): Iteration cycle
Iteration: While ||r(k−1)||2 > ε
1 k := k + 1.2 Find the most correlated direction jk := arg maxj /∈Sact
|r (k−1)j |.3 Update the set: S
(k)act := S
(k−1)act ∪ {jk}
A(k) := [A(k−1) ejk ], B(k) :=
[B(k−1) 0
0T 0
].
4 Compute current solution and residual:z(k) := arg minz ||A(k)z− y||22 + λ||B(k)z||22, r(k) = y− A(k)z(k).
Remark: The residual obtained at each cycle is strictly decreasing.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Reducing Complexity
As complexity at each step is O(n + k)3, k << n, there is aneed for cost reduction.
Initially, the matrix C (0) = [A(0)TA(0) + λB(0)] is inverted, inorder to solve the minimization task.
Similarly, at step k the matrix C (k) = [A(k)TA(k) + λB(k)] isinverted.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Reducing Complexity
However, as large part of the inverted matrix at each stepremains unchanged, we could avoid the inversion by applyingthe matrix inversion lemma (MIL):
C (k)−1=
[C (k−1)−1
+ 1tC
(k−1)−1αT
jkαjkC
(k−1)−1 −1tC
(k−1)−1αT
jk
−1tαjkC
(k−1)−1 1t
],
where αjk = eTjkA(k−1), t = 1−αjkC
(k−1)−1αT
jk.
Achieved computational cost: O(n + k)2 per step.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
sinc : 1-D case
−1 −0.5 0 0.5 1−40
−20
0
20
40Noisy data versus estimated data with KROMP
−1 −0.5 0 0.5 1−10
0
10
20Original data versus estimated data with MSE=0.05671
(a) KROMP with λ = 1.
−1 −0.5 0 0.5 1−40
−20
0
20
40Noisy data versus estimated data with ADM
−1 −0.5 0 0.5 1−10
0
10
20Original data versus ADM estimated data with MSE=0.09278
(b) ADM with λ = 1 and µ = 0.1.
Figure: Reconstruction of y = 20sinc(2πx) over the presence of 20dBGaussian noise and 10% outliers with values taken randomly in [−20, 20].(a) using KROMP and (b) using ADM (Giannakis-Mateos 2012). TheMSE is computed over 10000 data sets of 201 points.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
sinc : 2-D case
−10
1
−1−0.8−0.6−0.4−0.200.20.40.60.81−40
−20
0
20
40
60
Noisy data versus estimated data
−10
1
−1−0.8−0.6−0.4−0.200.20.40.60.81−20
0
20
40
60
Original data versus estimated data
Figure: Reconstruction of f (x , y) = 50sinc(π√x2 + y2) over the
presence of 15dB Gaussian noise and 10% outliers with values takenrandomly in [−20, 20]. KROMP attains a MSE = 1.022, while ADMattains a MSE = 2.205 (computed over 10000 data sets of21× 21 = 421 points).
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Application in image denoising: Lena
Original Image Noisy Image, PSNR=20.39dB
Figure: The original image of Lena (512× 512 pixels) and its noisycounterpart. Image corrupted with 20dB Gaussian noise and impulsenoise uniformly distributed with values ±100.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Application in image denoising: Lena
Restored Image with KROMP, PSNR=31.2dB Restored Image with ADM, PSNR=27.07dB
Figure: Left: reconstructed image using KROMP attains aPSNR = 31.2dB. Right: reconstructed image using ADM attains aPSNR = 27.07dB. In both algorithms parameters are set usingcross-correlation so that the MSE is minimized.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Application in image denoising: Boat
Original Image Noisy Image, PSNR=20.2dB
Figure: The original image of Boat (512× 512 pixels) and the noisyimage. Image corrupted with 20dB Gaussian noise and impulse noiseuniformly distributed with values ±100.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Application in image denoising: Boat
Restored Image with KROMP, PSNR=29.32dB Restored Image with ADM, PSNR=26.23dB
Figure: Left:reconstructed image using KROMP attains aPSNR = 29.32dB. Right:reconstructed image using ADM attains aPSNR = 26.23dB. In both algorithms parameters are set usingcross-correlation so that the MSE is minimized.
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Conclusions
An (OMP)-based algorithm is proposed for use in nonlinearestimation
The performance of the algorithm exceeds its predecessor(ADM) in numerous examples
The residual obtained at each cycle is proved to be strictlydecreasing
Recent applications in image-denoising seem to verify ourclaim towards better performance
The heavy computational cost is reduced using MIL; otherpopular techniques (e.g., Cholesky and QR factorization)seem to work also
(OMP) is a well studied algorithm, hence rich properties arethere to reinforce our work
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Future work
1 Construction of a more suitable stopping criterion
2 Complexity reduction
KROMP using QR decompositionKROMP using Cholesky factorization
3 Theoretical results for stability of the method
Conditions on the recovery of the exact support of the outliervectorBounds on the approximation of the solution
4 Applications of KROMP as an image or audio denoisingalgorithm
Codes and more at: http://bouboulis.mysch.gr/kernels.html
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)
Introduction Problem formulation Unravelling KROMP Experimental Results Conclusions and future work
Thank you
Thank you!Questions please
G. Papageorgiou, P. Bouboulis, S. Theodoridis DI, University of Athens, Greece
ROBUST KERNEL-BASED REGRESSION USING ORTHOGONAL MATCHING PURSUIT (OMP)