Recent thought: a "pattern" is "something a neural network can detect".
A "dog" is "something humans recognize as dog". Nothing more, nothing actually inherent to the animal. Most of what modern science does is redefine things to be more clear-cut: "dog is an animal that shares this much DNA with a dog prototype". But that's not *finding out* what a dog is, it's redefining.
(see also, whale is not a fish)
IOW, Plato was wrong. There is no ideal of dogness. What we see is what we get.
This is why artificial neural nets do so well: they work similar to humans. They can capture the nuance of dogness, because a neural net is the right structure to encode it. Because "dogness" is just that, a configuration of a neural net.
A computer *could* recognize a dog by the biological definition, but not from a photo. You could describe a dog by properties, but that's just describing what the neural net does. So why not go to the source.
