Why false?
Even if you fix all of the incorrect labels, that doesn’t mean your model will be able to predict all of them correctly.
but if I get rid of all Errors due to incorrectly labeled data — I will reduce overall dev set error to 11.2, isn’t it?
No, because even if the examples are labeled correctly, you don’t know if they will all be predicted correctly. Maybe the model isn’t very good, or isn’t complex enough for your data set.
Training almost never yields perfect predictions on any part of the data set.
1 Like
do you mean, that after fixing labels in next iteration (evaluation of model) result can be less than 15,3. But it will not stabely be reduced by 4.1% and can not give error around 11.2? Can be any number from 11.2 to 15,3?
I think I agree with you, but I’m not certain I understand your reply.
1 Like