Residual post-norm & cosine attention

How this idea evolved

Suggested from citations. No curator has recorded what this idea builds on yet. These ideas from the same theme come from papers that Swin V2 cites, directly or one step removed. Citation is a fact; the connection between the ideas is not verified.

• Shifted-window hierarchical ViTSwin · 2021cited directly

  Attention inside local windows that shift between layers, with multi-scale feature maps.

• Vision Transformer (ViT)ViT · 2020cited directly

  Split an image into patches and feed them to a plain Transformer as tokens.

• Convolution–attention hybridsCoAtNet · 2021cited directly

  Stack convolution stages before attention stages to get both inductive bias and capacity.