The assignation of the labels is what tells the network what you’re trying to teach it. Not the more granular features that constitute the reason that this particular input has the label and that one over there doesn’t, but the one-level-up-in-abstraction label. It isn’t the two eyes, the nose, the hair, the arms, the torso, the legs, the feet, the clothes. It’s the person that is made up of those things. If you then take a bunch of people and some have wry smiles and some don’t, and you label them as wry or not wry, then you’re telling the network what you’re looking for. But in that case it isn’t the look in the eye or the tilt of the head, the slightly pursed lips and the small furrow in the brow. It’s the wry smile made up of all those elements.
And here we arrive at the next phase in artificial intelligence: the phase guided by an individual’s taste. What I think is wry may not be what you think is wry. But that’s OK. What I think is good theater may not be what you think is good theater. But that’s OK. That’s called taste, discernment, opinion.
But this does raise the question: what are you teaching it? Do you know? Does it know? If the target is not very clear and very distinct, then it won’t work.