This idea relates to one of the key results in neural networks, where it has been shown that a sufficiently large single hidden layer network can approximate any function. (In essence a bunch of piece wise segments that can approximate the function along an interval.)