The softmax acts on the whole matrix
Isn’t the softmax applied vector-wise, thereby breaking the apparent transpose symmetry?
Isn’t the softmax applied vector-wise, thereby breaking the apparent transpose symmetry?