23. B.Lerner, H.Guterman, M. Aladjem , I.Dinstein
and Y.Romem, (1998) "On pattern classification with Sammon's nonlinear
mapping- An experimental study", Pattern Recognition , Vol.31, No
4, pp.371-381.
|
In [23] we discuss neural networks (NN’s) for multivariate
data projection, namely a NN implementation of the multidimensional scaling
(known in pattern recognition literature as Sammon’s mapping), auto-associative
NN for principal component analysis, and a multilayer perceptron (MLP)
when acting as a feature extractor. We compare their accuracy for discrimination
of the chromosome types “13”,”19”, and “x” using 100 patterns of each type.
The chromosome patterns were represented by 64 density profile features,
which are integral intensities along sections perpendicular to the medial
axis of the chromosome. The outputs of the NN paradigms are used to project
the patterns onto two-dimensional plots and to train and test an MLP classifier.
The two-dimensional projection maps are visually analyzed and the MLP probability
of correct classification evaluated for various numbers of extracted features.
For these NN paradigms applied to the chromosome analysis we found a strong
relationship between a highly visual exploratory projection map and high
discriminatory power. Superior paradigms for exploratory data projection
were found to be superior paradigms for discrimination and vice versa.
The MLP projection is found to be a preferred feature extraction paradigm
for both pattern classification and pattern visualization. Although originally
designed and used for exploratory data projection, Sammon’s mapping is
found to have impressive classification capability for chromosome data.
The auto-associative NN discloses a trade-off between an acceptable generalization
capability and a beneficial data compression.
|