Title: Frame, Reproducing Kernel and Learning
1Frame, Reproducing Kernel and Learning
- Alain Rakotomamonjy
- Stéphane Canu
Perception, Systèmes et Information Insa de
Rouen, 76801 St Etienne du Rouvray France
Alain.Rakoto,Stephane.Canu_at_insa-rouen.fr
http//asi.insa-rouen.fr/arakotom
2Motivations
- Wavelet-based approximation (wavelet or ridgelet
networks) are regularization networks? - Construction of multiresolution scheme of
approximation - kernel adapted to the structures of function to
be learned
3Motivations Ctd.
- Frame based framework for learning
Approximating highly oscillating structure
Without losing regularity in smooth region
4Road Map
- Introduction on Frame
- From Frame to Kernels
- From Frame kernels to learning
- Conclusions and perspectives
5Frame A definition
- H Hilbert Space dot product
A sequence of elements of H
is a frame of H if there exists A,B gt O s.t
A,B are the frame bounds
6Frame definition Ctd.
- Frame intepretation
- Frame allows stable representation
- as for all f in H
Frame "Basis" linear dependency redundancy
being a dual frame of Fn in H
7Particular cases of Frame
- Tight Frame
- Frame with bounds s.t AB
- Orthonormal Basis
- AB1
- Riesz Basis
- Frame elements are linearly independent
8Examples of Frame
- Tight Frame of IR2
-
-
-
- Frame of L2(IR)
F2
F1
F3
Y is an admissible wavelet
9Road Map
- Introduction on Frame
- From Frame to Kernels
- From Frame kernels to learning
- Conclusions and perspectives
10Frameable RKHS
- Condition for having a RKHS
Suppose H is a Hilbert space of function
and a frame of H
H is a RKHS if
On a frameable Hilbert Space, this is equivalent
to
The Reproducing Kernel is
11Construction of Frameable RKHS
- A Practical way to build a RKHS
- F is a Hilbert Space of function
A finite set of F elements such that
?
?
is a RKHS with Fn as frame elements
12Example of Frameable RKHS
- frameable RKHS included in L2(IR)
Fi L2 function (e.g Fi is a wavelet) span
Fii1N is a RKHS
Example
3 wavelets at same scale j
span a RKHS with kernel
13Road Map
- Introduction on Frame
- From Frame to Kernels
- From Frame kernels to learning
- Conclusions and perspectives
14Semiparametric Estimation
Learning from training set (xi,yi)i1..N
Semiparametric framework
One looks for the minimizer of the risk functional
in a space H spanYii1m H being a RKHS
Under general conditions,
spanYii1m parametric hypothesis space
15Semiparametric Estimation
- Parametric hyp. space is a frameable RKHS
P is a frameable RKHS spanned by Fn, with P ?
H, H RHKS
Semiparametric estimation on H with P as a
parametric hyp. space
One looks for the minimizer in H of
As spaces are orthogonal, backfitting is
sufficient for estimating f
16Semiparametric Estimation
- Frame view point
- H frameable
- H defined by kernel K
H P N
P Frameable RKHS, N Frameable RKHS
H
N "unknown component" to be regularized
P ? N due to linear dependency of frame
P "known component" not to regularized
KNKH-KP
P Frameable RKHS
17Multiscale approximation
H is splitted in different spaces Fii1m-1 and
H0
And any space Hi or Fi is a RKHS
Hi Trend Spaces
Fi Details Spaces
18Multiscale Approximation Ctd.
At each step j, trend obtained at step j-1 is
decomposed in trend and details
H
H2
F2
H1
F1
H0
F0
19Multiscale Approximation Ctd.
- Validity
- At each step, representer Theorem Hypothesis must
be verified - Solution
20Illustration on toy problem
Function to be learned
Data
xi N points from the random sampling of 0, 10
Algorithm
- SVM Regression
- Multiscale Regularisation networks on Frameable
RKHS
Sin/Sinc based kernel
Wavelet based kernel
21Results
- N902
- Results are averagerad over 300 experiments and
normalized with regards to SVM performance
Wavelet Kernel
Sinc Kernel
SVM
1 0.096
0.9297 0.312
0.5115 0.098
L2 error
0.7252 8.022
0.8280 0.025
1 0.028
22Plots of typical results
23Road Map
- Introduction on Frame
- From Frame to Kernels
- From Frame kernels to learning
- Conclusions and perspectives
24Summary
- new design of kernel based on frame elements
- algorithm for multiscale learning
- But
- no explicit definition of kernel
- Time-consuming
25Future work
- Multidimensional extension
- Tight Frame of multidimensional wavelet
- Using a priori knowledge on the learning problem
- How to choose the frame elements?
- Theoretical justification and analysis of
multiscale approximation
26(No Transcript)