 The other important unit is a feed-forward net. This is just a simple feed-forward neural network that is applied to every one of the attention vectors. These feed-forward nets are used in practice to transform the attention vectors into a form that is digestible by the next encoder block or decoder block.