Probably Approximately Correct

Universal Approximation Theorem

With what activation functions is learning possible?