Dpfp

We show the formal equivalence of linearised self-attention mechanisms and fast weight controllers from the early ’90s, where a slow neural net learns by gradient descent to program the fast weights of another net through sequences of elementary programming instructions which are...