A data transformation changes how data is represented while preserving information useful for a task. In machine learning, transformations help turn difficult patterns into forms that simpler rules can use. section 1.4
Recognizing handwritten digits
A handwritten-digit system might transform images into pixel histograms or counts of closed loops. These representations can help distinguish a 6 from an 8, but hand-written rules become brittle as new handwriting requires additional transformations and interacting exceptions. section 1.4
Common transformations include:
- Coordinate changes for two-dimensional classification. section 1.4
- Linear projections, translations, and nonlinear selections such as retaining points where x > 0. section 1.4
- Automatically searched operations selected by feedback on a development dataset. section 1.4
Figure 1.4 Coordinate change
Compare the original representation with a coordinate change that can make classification easier. section 1.4
