Scribd
Upload
118 views

4.2 4.2 Convex Sets, Convex Hulls and Linear Separability: X X C C X X

The document defines and discusses convex sets and convex hulls. It defines a convex set as one where any line segment between two points in the set is also contained in the set. Convex hulls are defined as the smallest convex set containing a given pattern set. The document notes that if the convex hulls of two pattern sets are disjoint, the sets are linearly separable by a hyperplane. It provides an example of iris data where the convex hulls of two classes are separable by a hyperplane, while two other classes are not separable.

Uploaded by

zym1003
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
118 views

4.2 4.2 Convex Sets, Convex Hulls and Linear Separability: X X C C X X

The document defines and discusses convex sets and convex hulls. It defines a convex set as one where any line segment between two points in the set is also contained in the set. Convex hulls are defined as the smallest convex set containing a given pattern set. The document notes that if the convex hulls of two pattern sets are disjoint, the sets are linearly separable by a hyperplane. It provides an example of iris data where the convex hulls of two classes are separable by a hyperplane, while two other classes are not separable.

Uploaded by

zym1003
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

4.

2 Samples November 17, 2004 13:3

4.2
4.2 Convex Sets, Convex Hulls and Linear Separability

Consider two pattern sets X1 and X2 sampled from two classes C1 and C2 . Recall that a vector of input fea-
Many subsets of Rn contain the pattern sets X1 and X2 . In the present context tures is called a pattern. The
vector space from which valid
we are interested only in the smallest convex sets that contain these sets, patterns may be derived is then
and very shortly we will see why. But rst the following denitions are in the pattern space, typically Rn .
order. In general any ndimensional
pattern can be represented by
a point in pattern space, Rn .
Denition 4.2.1

Let X, Y S Rn , then S is convex iff X + (1 )Y S, 0


1, X, Y S. Equivalently, a set S is convex if it contains all points on all line
segments with end points in S.

In Rn , spheres, ellipsoids and cubes are examples of convex sets. Tori


and disconnected sets are not convex. Figure 4.1 portrays some convex and
non-convex sets.

Fig. 4.2 Convex and non-convex sets in R2

Denition 4.2.2

The convex hull, C(Xi ), of a pattern set Xi is the smallest convex set in Rn which
contains the set Xi . Equivalently, consider every convex set S , such that X
i S
Rn , I, where I is an index set. Then the convex hull of Xi , C(Xi ) = I S .

This only means that we need to take the intersection of all Rn subsets
that contain the pattern set in question. This intersection yields the smallest
convex set in Rn that contains the pattern set. Figure 4.2 shows a computer
4.2 Samples November 17, 2004 13:3

2 Neural Networks: A Classroom Approach

generated convex hull of the Iris data [12], using MATLAB. Notice how
the convex hull of iris sestosa is disjoint from the convex hulls of iris
versicolor and iris virginica. However, the convex hulls of iris versicolor
and iris virginica are not separable.
The problem of convex hull com-
putation is an interesting area of
research. Neural networks have
been applied to this problem
(see [432]).

Fig. 4.3 MATLAB generated convex hulls for Iris data [12]

Note: This is a view of the 4-


It should be intuitively clear that if the convex hulls of two pattern sets
dimensional Iris data projected
on to the xi3 xi4 axes. The con- are non-overlapping, we can dene a separating hyperplane that slices the
vex hull was generated using pattern space into two halves, such that the pattern sets are separated. This
MATLABs convhull command. leads to the concept of linear separability.

Denition 4.2.3

Two pattern sets Xi and Xj are said to be linearly separable if their convex hulls
are disjoint, that is if C(Xi ) C(Xj ) = .

As shown in Fig. 4.3 since some minimum distance separates the two
convex hulls, there exists a hyperplane , that separates the two sets. In
the gures, we use the shorthand Ci to denote the convex hull C(Xi ). The
convex hull Ci is representative of the pattern class Ci .
4.2 Samples November 17, 2004 13:3

One separating hyperplane is


the perpendicular bisector of the
straight line joining the closest
two points on the disjoint convex
hulls.

Fig. 4.4 Linearly separable pattern classes and a separating hyperplane

You might also like