Explaining Explainability: Understanding Concept Activation Vectors