Comment by braised_babbage

No, neural networks cannot "represent any arbitrary function". Find a theorem that you think states otherwise, and then read what the theorem actually states.