Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
42 views

Machine Learning Course - Kernel Regression

This document provides an overview of kernel ridge regression and the kernel trick. [1] It establishes the relationship between the primal solution w? and the dual solution α? for ridge regression. [2] It introduces the representer theorem, which allows solving problems involving w? equivalently in terms of α?. [3] This paves the way for kernelizing problems by defining kernels based on inner products in feature spaces, without needing to explicitly specify the feature maps.

Uploaded by

nagybaly
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
42 views

Machine Learning Course - Kernel Regression

This document provides an overview of kernel ridge regression and the kernel trick. [1] It establishes the relationship between the primal solution w? and the dual solution α? for ridge regression. [2] It introduces the representer theorem, which allows solving problems involving w? equivalently in terms of α?. [3] This paves the way for kernelizing problems by defining kernels based on inner products in feature spaces, without needing to explicitly specify the feature maps.

Uploaded by

nagybaly
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Machine Learning Course - CS-433

Kernel Ridge Regression and


the Kernel Trick

Oct 31, 2019

changes by Martin Jaggi 2019, changes by Rüdiger Urbanke 2018, changes by Martin Jaggi
2016, 2017 Mohammad
c Emtiyaz Khan 2015
Last updated on: October 31, 2019
Motivation
The ridge solution w? ∈ RD has a
counterpart α? ∈ RN . Using dual-
ity, we will establish a relationship
between w? and α? which leads the
way to kernels.

Ridge regression
Recall the ridge regression problem

min 1
2 ky − Xwk2 + λ2 kwk2
w

For its solution, we have that

w? = (X>X + λID )−1X>y


= X>(XX> + λIN )−1y =: X>α?,

where α? := (XX> + λIN )−1y.

This can be proved using the follow-


ing identity: let P be an N × D ma-
trix while Q be D × N , and let both
PQ + I and QP + I be invertible.

(PQ + IN )−1P = P(QP + ID )−1

What are the computational com-


plexities for the above two ways of
computing w??
With this, we know that w? =
X>α? lies in the column space
of X>,
 
x11 x12 . . . x1D
 x21 x22 . . . x2D 
where X =  .. .. . . . ..  
xN 1 xD2 . . . xN D

The representer theorem


The representer theorem generalizes
this result: for a w? minimizing the
following function for any Ln,
N
X
min Ln(x> λ
n w, yn ) + 2 kwk
2
w
n=1

there exists α? such that


w ? = X> α? .

Such a general statement was orig-


inally proved by Schölkopf, Her-
brich and Smola (2001).
Kernelized ridge regression
The representer theorem allows us
to write an equivalent optimization
problem in terms of α. For exam-
ple, for ridge regression, the follow-
ing two problems are equivalent:

w? = arg min 1
2 ky − Xwk2 + λ2 kwk2
w
1 >
α = arg max − α (XX> + λIN )α + α>y
?
α 2
i.e. they both have the same
optimal value. Also, we can always
have the correspondence mapping
w = X>α.

Most importantly, the second


problem is expressed in terms of
the matrix XX>. This is our first
example of a kernel matrix.

Note: To see the equivalence, you can show


that we obtain equal optimal values for the
two problems. Take the gradient of the sec-
ond expression, to get (XX> + λIN )α − y.
Setting this to 0 and solving for α results
in α? = (XX> + λIN )−1y.

If we combine this with the representer the-


orem w? = X>α? we find back the dual
solution.
Advantages of kernelized ridge regression
First, it might be computationally
efficient in some cases when solving
the system of equations.

Second, by defining K = XX>, we


can work directly with K and never
have to worry about X. This is the
kernel trick.

Third, working with α is some-


times advantageous, and provides
additional information for each dat-
apoint (e.g. as in SVMs).

Kernel functions
The linear kernel is defined below:
 >
x1 x1 x> >

1 x2 . . . x1 xN
>
 x> > >
2 x1 x2 x2 . . . x2 xN

K = XX =  .  .. ... .. .
. 
x> > >
N x1 xN x2 . . . xN xN

Kernel with basis functions φ(x)


with K := ΦΦ> is shown below:
> > >
 
φ(x1) φ(x1) φ(x1) φ(x2) . . . φ(x1) φ(xN )
 φ(x2)>φ(x1) φ(x2)>φ(x2) . . . φ(x2)>φ(xN ) 
. . . .
. ..

 . . . 
φ(xN )>φ(x1) φ(xN )>φ(x2) . . . φ(xN )>φ(xN )
The kernel trick
A big advantage of using kernels
is that we do not need to specify
φ(x) explicitly, since we can work
directly with K.

We will use a kernel function


κ(x, x0) and compute the (i, j)-th
entry of K as Kij = κ(xi, xj ). A
kernel function κ is usually associ-
ated with a feature map φ, such that

κ(x, x0) := φ(x)>φ(x0) .

For example, for the trivial or ‘lin-


ear kernel’ κ(x, x0) := x>x0, the fea-
ture map is just the original features,
φ(x0) = x0.

Another example: For data x ∈


R3, the kernel κ(x, x0) := (x>x0)2 =
(x1x01 +x2x02 +x3x03)2 corresponds to
>
 2 2 2 √ √ √ 
φ(x) = x1, x2, x3, 2x1x2, 2x1x3, 2x2x3

The good news is that the evalua-


tion of a kernel is often faster when
using κ instead of φ.
Visualization
Why would we want such general
feature maps?

See video explaining linear sepa-


ration in the kernel space (where
φ(x) maps to) corresponding to
non-linear separation in the original
x-space: https://www.youtube.
com/watch?v=3liCbRZPrZA

Examples of kernels
The above kernel is an example of
the polynomial kernel. Another ex-
ample is the Radial Basis Function
(RBF) kernel.
 
1
κ(x, x ) = exp − (x − x0)>(x − x0)
0
2
See more examples in Section 14.2
of Murphy’s book.

A natural question is the following:


how can we ensure that there exists
a φ corresponding to a given ker-
nel K? The answer is: as long as the
kernel satisfies certain properties.
Properties of a kernel
A kernel function must be an inner-
product in some feature space. Here
are a few properties that ensure it is
the case.
1. K should be symmetric, i.e.
κ(x, x0) = κ(x0, x).
2. For any arbitrary input set
{xn} and all N , K should be
positive semi-definite.
An important subclass is the
positive-definite kernel functions,
giving rise to infinite-dimensional
feature spaces.
Exercises
1. Understand the relationship w? = X>α?. Understand the state-
ment of the representer theorem.
2. Show that the primal and dual formulations of ridge regression
are equivalent. Hint: show that the optimization problems corre-
sponding to w and α have the same optimal value.
3. Get familiar with various examples of kernels. See Section 6.2 of
Bishop on examples of kernel construction. Read Section 14.2 of
Murphy’s book for examples of kernels.
4. Revise and understand the difference between positive-definite
and positive-semi-definite matrices.
5. If you are interested in more details about kernels, read about Mer-
cer and Matern kernels from Kevin Murphy’s Section 14.2. There
is also a small note by Matthias Seeger on the git repository under
lectures/07, in particular for the case of infinite dimensional φ.

You might also like