MUREL: Multimodal Relational Reasoning for Visual Question Answering