Information Technology Reference
In-Depth Information
2.5
Structural Classifier
TDNN Classifier
Neural Abstraction Pyramid
TDNN+Structural
Pyramid+TDNN+Struct
2
1.5
1
0.5
0
0
1
2
3
4
5
6
7
8
9
(a)
% Reject
2.5
Structural Classifier
TDNN Classifier
Neural Abstraction Pyramid
TDNN+Structural
Pyramid+TDNN+Struct
2
1.5
1
0.5
0
0
1
2
3
4
5
6
7
8
9
(b)
% Reject
Fig. 5.7.
Learning a hierarchy of sparse features - performance of different digit classifiers:
(a) test set; (b) validation set.
best stimuli for the feature detectors, one can see that these are not similar in terms of
a simple pixel-based distance measure, but in terms of their recursive decomposition
to substructures. Hence, the pyramidal digit representation becomes increasingly
invariant against distortions when going up in the hierarchy.
The extracted features facilitate recognition of the digits. When used as input to
an FFNN-classifier, the recognition performance observed was very satisfactory. It
outperforms any single classifier that has been tested on that dataset and is about
as good as the combination of the TDNN and the structural digit recognizer. When
combined with these two classifiers, the recognition performance improves further.
Search WWH ::
Custom Search