Training for better than human performance

In the course “Structuring Machine Learning Projects” it mentions that a challenge to perform better than humans is to get the data, since humans are unable to perform that well. However, it seems to me that in many cases you can just augment data. E.g., in image recognition, you could take pictures where humans are 100% certain, and then blurry them, lower the resolution, etc, etc, to a point that for a human is too dificult to recognize. Would something like that work or am I missing something?

Augmenting a data set in order to increase the number of labeled examples is a common method.

1 Like