Problem with that is the "errors" in the real output bins will be encoded into the next generation of neural net and no progress will be made. On the other hand

Problem with that is the "errors" in the real output bins will be encoded into the next generation of neural net and no progress will be made. On the other hand, the dataset could be used as a starting point that could be one-time filtered and improved by a human.