arXiv:2110.02178v1 [cs.CV] 5 Oct 2021. Technical report N 3 P Flatten image patches Linear N d Positional encoding Transformer L Linear Logits (a) Standard visual transformer (ViT) MobileViT block Local representations Transformers as Convolutions (global representations) Fusion W H C X Conv-nn? y Conv-1x1 W H d P N d Transformer L P d W H d

2011