Subsections

• 15.7.1 Optical Character Recognition (OCR)
• 15.7.2 Text Categorization and Text Mining
• 15.7.3 Active Learning in Drug Design
• 15.7.4 Other Applications

15.7 Applications

15.7.1 Optical Character Recognition (OCR)

One of the first real-world experiments carried out with SVM was done at AT&T Bell Labs using optical character recognition (OCR) data ([17,54]). These early experiments already showed the astonishingly high accuracies for SVMs which was on a par with convolutive multi-layer perceptrons. Below we record the classification performance of SVM, some variants not discussed in this chapter, and other classifiers on the USPS (US-Postal Service) benchmark (parts from [62]). The task is to classify handwritten digits into one of ten classes. Standard SVMs as presented here already exhibit a performance similar to other methods.

Table 15.2: Classification results on the USPS data set

Linear PCA & Linear SVM ([60])	$8.7\,{\%}$
k-Nearest Neighbor	$5.7\,{\%}$
LeNet1 ([13])	$4.2\,{\%}$
Regularised RBF Networks ([50])	$4.1\,{\%}$
Kernel-PCA & linear SVM ([60])	$4.0\,{\%}$
SVM ([54])	$4.0\,{\%}$
Invariant SVM ([57])	$3.0\,{\%}$
Boosting ([22])	$2.6\,{\%}$
Tangent Distance ([62])	$2.5\,{\%}$
Human error rate	$2.5\,{\%}$

A benchmark problem larger than the USPS data set ($7291$ patterns) was collected by NIST and contains $120{,}000$ handwritten digits. Invariant SVMs achieve an error rate of $0.7\,{\%}$ ([19]) on this challenging and more realistic data set, slightly better than the tangent distance method ($1.1\,{\%}$) or single neural networks (LeNet $5$: $0.9\,{\%}$). An ensemble of LeNet $4$ networks that was trained on a vast number of artificially generated patterns (using invariance transformations) is on a par with the best SVM and also achieved $0.7\,{\%}$ ([36]).

15.7.2 Text Categorization and Text Mining

The high dimensional problem of text categorization seems to be an application for which SVMs have performed particularly well. A popular benchmark is the Reuters-$22{,}173$ text corpus, where Reuters collected $21{,}450$ news stories from 1997, and partitioned and indexed them into $135$ different categories to simplify the access. The feature typically used to classify Reuters documents are $10{,}000$-dimensional vectors containing word frequencies within a document. With such a coding SVMs have achieved excellent results (see e.g. [31]).

15.7.3 Active Learning in Drug Design

A challenging problem of computational chemistry is the discovery of active chemical compounds. The goal is to found the compounds that bind to a certain molecule in as few iterations of biological testing as possible. An astonishingly effective application of SVM to this problem has been recently proposed by [75].

15.7.4 Other Applications

There are numerous other applications to which SVM were successfully applied. Examples are object and face recognition tasks ([46]), inverse problems ([71,77]), gene array expression monitoring ([15]), remote protein homology detection ([30]), or splice site detection ([79,64]). This list can be extended.