"Animal" is an arbitrary definition that exists in one's head. Unless you assume an idealist position, the only way it can be defined is by a function that runs in one's head. It might be a very complicated function - sufficiently complicated that we can't really articulate it and resort to "I know it when I see it", and possibly self-modifying - but the same person will still produce the same output consistently when given the same input.
With NNs, fns are approximated by compressing a sample of points along an empirical dataset which stands in for the fn. In most cases, however, the fn doesnt exist.
There is no fn from "Image->Animal". The world is ambiguous, functions do not describe it. The different `x`s genuinely correspond to the same `y`, here: the same image (occluded animal) could either be a dog or cat. There is no function to approximate.
Even when there is, compressing empirical samples along a trivial interval is a terribly fragile method: consider sampling the addition function (x,y)->x+y over anything less than the infinite domains of its inputs.
Starting with an empirical distribution of "the solved problem" will never build an intelligent system. Intelligence is what animals have because the world is ambiguous and not already ready-to-hand.
Intelligence is what you do when there isnt "functions lying around" to approximate. I doubt almost anything in CSci is even relevant to solving this problem. It's largely a bioengineering one. Being the agent in the world which manufactures functions by disambiguating it.