Open Models Almanac
Methodsmodel architecturetoken mixer

single mention · model architecture

Windowed multi-head attention pooling

Within each patch window, multi-head attention pools four patches into one vector using their mean as the query.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See token mixer for the methods filed beside it.

Also called Attention pooling in windows.

Filed under model architecture › token mixer