[ML] Feature scaling

Purpose:
Make the gradient descent finding the global minimum faster.

Idea:
    Make sure features are on a similar scale.
   
    e.g
    make the feature scale unified as 0 <= x <=1
    i.e
    if a scale is of 0 ~ 2000(max),
    divide it by 2000(range, means max - min) , then we end up with the value
    between 0 and 1.
   
    This will make the graph of gradient descent as close to a circle shape, which makes
    the finding of global minimum much more easier.
   
    Thus, we want every features into *approximately* a
    -1 <= Xi <= 1 range.
   
    X0 is 1 by define.
   
Mean normalization:
    i.e
    make every input feature minus the mean of all features.
    => X1 = (input - mean) / 2000( range of the scale, OR the standard deviation)
   
    This will make the value more close to 0.