Extracting Formulae in Many-Valued Logic from Deep Neural Networks