Surpassing human-level performance Q

I think it must be important to think about the question of the accuracy of labeling when it comes to this question. If we rely on humans to do the labelling, then there must be an error associated with the labelling. Then we can’t expect a computer to surpass this, and improvement must be overfitting. Please can you clarify and how can we estimate the labeling uncertainty?

The issue of mislabelled data was answered later in the course.