Learning Normalitzation

Image normalization:

1. We subtract a fixed average of each channel (centralize)

Note: This average was computed by a sample image a long time ago, we should update it with the new lighting

2. We divide by 255

Label normalization

1. We compress it to the output shape
2. We set all values over 20 (abs val) to +-20

Note: Should we divide by 20 after?

Note: Should we use the tan() of the value?