I think the "static is a cheetah" example just highlights that the neural net is not identifying the best features with which to identify a cheetah. Or alternat

I think the "static is a cheetah" example just highlights that the neural net is not identifying the best features with which to identify a cheetah. Or alternatively, if, during training, the neural net was only fed pictures of nature with or without cheetahs in them, then what it's really telling you is not the probability that a picture contains a cheetah, but rather the conditional probability that a picture contains a cheetah given that it is a picture of nature. In other words, that picture of static is most likely well outside the domain of the training set, so classifying it involves a large extrapolation, with all the attendant amplification of errors. Perhaps what we need is a classifier that can tell when a picture is significantly outside of its training experience and say "I've never seen anything like that before" instead of giving an arbitrary classification.