So he's going to apply a model trained on... well, basically crap, to "fix" the crap so it is somehow crap no longer, and then retrain the model on the new, re-crappified crap. Sounds like a tremendous waste of energy.