ch2-general-system-design.tex 2.2 KB

123456789101112131415161718192021222324252627282930313233343536
  1. %!TEX root = write-math-ba-paper.tex
  2. \section{General System Design}
  3. The following steps are used for symbol classification:\nobreak
  4. \begin{enumerate}
  5. \item \textbf{Preprocessing}: Recorded data is never perfect. Devices have
  6. errors and people make mistakes while using the devices. To tackle
  7. these problems there are preprocessing algorithms to clean the data.
  8. The preprocessing algorithms can also remove unnecessary variations
  9. of the data that do not help in the classification process, but hide
  10. what is important. Having slightly different sizes of the same symbol
  11. is an example of such a variation. Four preprocessing algorithms that
  12. clean or normalize recordings are explained in
  13. \cref{sec:preprocessing}.
  14. \item \textbf{Data multiplication}: Learning systems need lots of data
  15. to learn internal parameters. If there is not enough data available,
  16. domain knowledge can be considered to create new artificial data from
  17. the original data. In the domain of on-line handwriting recognition,
  18. data can be multiplied by adding rotated variants.
  19. \item \textbf{Feature extraction}: A feature is high-level information
  20. derived from the raw data after preprocessing. Some systems like
  21. Detexify take the result of the preprocessing step, but many compute
  22. new features. Those features can be designed by a human engineer or
  23. learned. Non-raw data features have the advantage that less
  24. training data is needed since the developer uses knowledge about
  25. handwriting to compute highly discriminative features. Various
  26. features are explained in \cref{sec:features}.
  27. \end{enumerate}
  28. After these steps, it is a classification task for which the classifier has to
  29. learn internal parameters before it can classify new recordings.We classified
  30. recordings by computing constant-sized feature vectors and using
  31. \glspl{MLP}. There are many ways to adjust \glspl{MLP} (number of neurons and
  32. layers, activation functions) and their training (learning rate, momentum,
  33. error function). Some of them are described in~\cref{sec:mlp-training} and the
  34. evaluation results are presented in \cref{ch:Optimization-of-System-Design}.