Professional Documents
Culture Documents
Prepared By :
Motivation
1 l
C xi xi
V CV
T
l i 1
Kernel PCA
Since PCA won’t work for non-linear
variations, map the data to an
approximately linear space F
: F N
Kernel PCA
New problem:
1 l
C ( xi ) ( xi ) T (1)
l i 1
V C V
Solving this directly is expensive
Instead, use the ‘kernel trick’
Kernel PCA
Eigenvectors will lie in the span of projected
data ( x ),..., ( x )
1 l
K ij : ( ( xi ) ( x j ))
T
Kernel PCA
Substituting (1) and (3) into (2), we get:
l
1 l l
( ( xi ) i ( xi )) ( ( xi ) ( xi ) ( xi ) i ( xi )), k 1,..., l
T
i 1 l i 1 i 1
l K
The trick to Kernal PCA
We can easily have our subspace, but how to
construct K?
Consider a mapping with polynomial degree 5,
and 16x16 pixel images
It would require 1010 dimensionality!
Instead, just define a kernel function
k ( x, y ) ( ( x) ( y ))
T
Kernel function
Using the kernel function to build the kernel
matrix will make computation reasonable
Arandjelovic and Cipolla used
.6 ( x y ) T ( x y )
k ( x, y ) e
found empirically
Comparing datasets using
RAD
Resister-Average Distance
Based on Kullback-Leiber divergence
Datasets represented as probability distributions
1 1 1
DRAD ( p, q) ( DKL ( p || q ) DKL (q || p ) )
Kullback-Leiber
divergence
p ( x)
Defined as DKL ( p || q) p( x) log 2 dx
q( x)
2 |C |
p