Feature scaling
lastmod 2018-08-17created 2015-12-302 edits
Feature scaling is a methodology for standardizing the range of features (machine learning) or independent variables of data. In data processing it is defined as data normalization, a processing task for a previous stage.
In the Machine learning
If data sizes differ greatly between features, the contour lines in Gradient Descent are closely spaced, so it can take a long time to find the Global optima. So you can divide Feature values by m or scaliing them to between -1 and 1. Put the other way around, using Feature scaling lets Gradient descent find the result faster.
You can also use Mean normalization, which subtracts the mean from every feature to make the mean 0.
A more common method is to apply mean normalization and then divide by max-min or the standard deviation.