[dguest]

Cases where data augmentation works always pose a fun challenge: clearly if there's some way to manipulate the data which shouldn't alter the algorithm output, there's also some way to make the ML invariant to that manipulation. If you can create an invariant algorithm then the augmentation should be unnecessary.

There is some interesting work on making ML invariant to rotations and such, but I'd be curious if there's an algorithm which is invariant to this. I could imagine that convolutions and pooling might be relatively invariant to this technique, for example, as long as the algorithm isn't doing a lot with the large scale structure of the image.