From "Where" to "What": Towards Human-Understandable Explanations through Concept Relevance Propagation