Eager to Stop: Efficient Falsification of Deep Neural Networks