selected publications blog posting ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision 2021