Information Theory and Principal Component Analysis - Biomedical Signal Analysis: Contemporary Methods and Applications - page 96

5

3.5 x 10

3

2.5

2

1.5

1

0.5

0 1 0

10
1 2 3

10 10 10 10

eigenvalue #

(a) digits data set (b) digits data set

(c) PCA result

Figure 3.6
NIST digits data set. In (a), we show a few samples of the 1000 28x28 gray-scale
pictures of the digits 2 and 4 used in the analysis. (b) shows the eigenvalue
distribution of the covariance matrix i.e. the power of each principal component,
and (c) a projection onto the first two principal components. At each
two-dimensional location, the corresponding picture is plotted. Clearly, the first two
PCs already capture the differences between the two digits.