They can model only continuous functions, more specifically any continuous function on compact subsets of ℝⁿ. They can approximate functions to an arbitrary level of accuracy, given sufficient neurons