MUTAN: Multimodal Tucker Fusion for Visual Question Answering