Logical Explanations for Deep Relational Machines Using Relevance Information