##### ###### ##### ### # # ### # # ###### ## ## ## ## ## ## ## # # # # # ## ##### #### ##### # # # # # # # #### ## # ## ## ## ## # # # # # ## ## # ###### ## ### # ### # ######
##### ###### ##### ### # # ### # # ###### ## ## ## ## ## ## ## # # # # # ## ##### #### ##### # # # # # # # #### ## # ## ## ## ## # # # # # ## ## # ###### ## ### # ### # ######
After computing attention separately in each head, the results are concatenated and projected back to the model dimension.
MultiHead(Q, K, V) = Concat(head_1, ..., head_h) * W_O
The output projection W_O learns to combine information from all heads. This final linear layer mixes the different attention patterns into a unified representation.
