Comment by s1t5

In one of his fastai videos Jeremy Howard makes the point that wrong labels can act as regularization and you shouldn't worry too much about them. I'm a bit skeptical as to how far you can push this but you certainly don't need perfect labelling.