MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
78<br />
As in the original space, in feature space, the correlation matrix C φ<br />
can be diagonalized and we<br />
have now to find the eigenvalues λi ≥ 0, i= 1... M,<br />
satisfying:<br />
Cv<br />
i<br />
φ<br />
= λ v<br />
i<br />
i<br />
( ) φ<br />
λi<br />
φ( )<br />
j i j i<br />
⇒ φ x , C v = x , v , ∀ i, j = 1,... M<br />
(5.10)<br />
All solutions v with λ ≠ 0 lie in the span of the<br />
Developping the left hand-side<br />
M<br />
j i j i l<br />
( x ), Cφv = ( x ),<br />
Cφ∑<br />
l ( x )<br />
φ φ α φ<br />
=<br />
M<br />
∑<br />
l=<br />
1<br />
1<br />
=<br />
M<br />
1<br />
=<br />
M<br />
l=<br />
1<br />
( x ),<br />
Cφ<br />
( x )<br />
α φ φ<br />
i j l<br />
l<br />
M<br />
∑<br />
l=<br />
1<br />
( x ),<br />
FF ( x )<br />
α φ φ<br />
i j T l<br />
l<br />
M<br />
( x ), ∑ ( x ) ( x ),<br />
( x )<br />
M<br />
i j j j l<br />
∑αl<br />
φ φ φ φ<br />
l= 1 j=<br />
1<br />
φ<br />
1<br />
M<br />
(x ),…, φ(x )<br />
( )<br />
M<br />
i i i j<br />
i i j<br />
j=<br />
1<br />
and we can thus write:<br />
Cv<br />
φ<br />
= λv = λ∑ α φ x<br />
(5.11)<br />
Replacing the latter expression in the definition of the correlation matrix, one gets:<br />
1<br />
M α φ φ φ φ λ α φ φ<br />
M M M<br />
i j j j l i i j<br />
∑ l ( x ), ∑ ( x ) ( x ), ( x ) =<br />
i∑ j ( x ), ( x ) . (5.12)<br />
l= 1 j= 1 j=<br />
1<br />
Using the kernel trick, one can define the Gram Matrix K , whose elements are composed of the<br />
dot product between each pair of datapoints projected in feature space, i.e.<br />
i j<br />
( ) φ( )<br />
K φ x , x .<br />
ij<br />
= Beware that K is M × M , where M is the number of data points.<br />
We can finally rewrite the expression given in (5.12) as an eigenvalue problem of the form:<br />
2 i<br />
i<br />
K α = Mλi<br />
Kα<br />
, i=<br />
1... M<br />
i<br />
i<br />
Kα<br />
= Mλα<br />
i<br />
(5.13)<br />
This is the dual eigenvalue problem of finding the eigenvectors<br />
i<br />
v of C.<br />
© A.G.Billard 2004 – Last Update March 2011