I have a scenario I have been wondering about.
If I save weights for a trained data set with 1 million things (images, audio, text, whatever) and then add another 1 million things and train further. Will I get similar results if I took 2 million things and trained a model? Would either perform better?
What I am trying to understand is how would you handle a model for something like anti-spam where it is updated around the clock but is also used to predict around the clock.