We introduce a geometric framework for deriving attention architectures. A sequence forms a memory $m\in M_n$, and an attention pairing $ρ_n\colon M_n\to\operatorname{Hom}(A,A)$ determines how that memory acts on each token state. Different representations of this action recover familiar architectures: softmax attention reads individual key--value pairs, linear attention accumulates their contributions in matrix statistics, and Intention fits a regression map from keys to values. The framework also suggests new architectures by changing the geometry of the query--key pairing, the function carried by each value, or the space representing the memory. ImageNet-1K experiments illustrate these constructions through geometric CAFormer variants, a ViT with query-dependent values, and a linear ViT with a wider memory representation.
翻译:暂无翻译