Here’s how: prior to the transformer, what you had was essentially a set of weighted inputs. You had LSTMs (long short term memory networks) to enhance backpropagation – but there were still some ...