A type of recurrent neural network that involves duplicating the first recurrent layer in the network so that there are now two layers side-by-side, then providing input sequences in their original order and their reversed order to each of the two layers.