@behohippy @saint Instead of timestep by timestep sequence modeling the attention allows us to pass sequential model in a parallel NN just like fully connected one, where the positional encoding helps us to know the sequence of each and we can remove the keys having less attention value...