【多模态论文解读】Align before Fuse: Vision and Language Representation Learning with Momentum Distillation