We need to worry about data balancing

[sg-s]

If we keep adding noise to the training data, then networks can "cheat" by simply marking everything as noise.

So every time a point is added to the training data, we need to make sure that the data is 'balanced", and no one label predominates.