"can't distinguish a corgi from a loaf of bread or specially designed white noise" You are talking about adversarial examples, which are in most cases extremely

"can't distinguish a corgi from a loaf of bread or specially designed white noise" You are talking about adversarial examples, which are in most cases extremely rare and not detrimental to the typical function of the image recognition system. Even humans have problems identifying some of these adversarial examples. The white noise is particularly troublesome from a technical standpoint. It goes to show that these learned manifolds are not perceptual manifolds.