Improved Fusion of Visual and Language Representations by Dense Symmetric Co-Attention for VQA 阅读笔记