01.11.2014 Views

MACHINE LEARNING TECHNIQUES - LASA

MACHINE LEARNING TECHNIQUES - LASA

MACHINE LEARNING TECHNIQUES - LASA

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

73<br />

5 Kernel Methods<br />

These lecture notes have sofar covered linear methods for performing a variety of computation,<br />

ranging from dimensionality reduction, clustering and classification to regression. The term linear<br />

refers to the assumptions that all of these transformations could be expressed as linear<br />

transformation of the form y= Ax (where y and xare given and A is the unknown<br />

transformation).<br />

Kernel methods relax the assumption of a linear transformation, so as to perform non-linear<br />

regression, classification, etc. Kernel methods proceed by first projecting the data through a nonlinear<br />

transformation into a feature space and then perform the same type of computation (e.g.<br />

classification or regression), as in the linear case, in the feature space. This principle is illustrated<br />

in Figure 5-1.<br />

Figure 5-1: Illustration of the principle of non-linear classification through Support Vector Machine. The data<br />

are first projected from input space to a feature space via a non-linear map φ . Linear separation is then<br />

performed in the feature space. Figure from "Learning with Kernels" by B. Scholkopf and A. Smola, MIT<br />

Press 2002.<br />

In these lecture notes, we will cover solely a subset of the existing Kernel methods, focusing on<br />

kernel methods that extend the algorithms see previously for dimensionality reduction (PCA, ICA,<br />

CCA to kernel PCA, kernel ICA, kernel CCA), for clustering (from K-Means to kernel K-means),<br />

for classification (from linear classifier to support vector machine) and for regression (from<br />

probabilistic regression to Gaussian Process Regression).<br />

5.1 The kernel trick<br />

All kernel methods we will see here are based on the so-called kernel trick. The kernel trick stems<br />

from a variety of fundamental mathematical principles which we will skip here, focusing solely on<br />

explaining its basic principle.<br />

© A.G.Billard 2004 – Last Update March 2011

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!