One of the most archetypal failure modes of current NN-based classifiers is, for example, happily reporting seeing a spotted feline given a picture of a couch with a leopard pattern.
Well, if I'm given the task to enumerate what I see in that image maybe I would also add the feline label. But I would add the picture_of attribute, if that's a thing.
I don’t think today’s classifiers are anywhere sophisticated enough to have a concept of something vs. a picture of something, a type of use–mention distinction, or ceci n’est un pipe if you will. The point was that people think they trained their classifier to recognize leopards, but in actuality it just learned to recognize the leopard coat pattern.
Yes, maybe. But it is at least unclear whether is inherently CNN or not (many people would vote for no). ImageNet training data is heavily texture dependent, therefore, the trained model leaning towards texture information more. Some recent papers did style transfer to randomize textures in order to train the model to focus more on the shape, the reported results seem good.
Exactly. I would also add (although without strong evidence to support my assertion) that it seems to be more a function of the training data than of the network architecture.