Multilayer RNNs (also called deep RNNs) is another concept. The idea here is that we add additional RNNs on top of the source network, where each added RNN is a different layer. The output of the hidden state of the first (or lowest) RNN is the input for the RNN of the next layer above it. The general prediction is usually calculated from the latent state of the most recent (highest current output) layer:

The preceding diagram shows a multilayer unidirectional RNN, where the layer number is indicated as superscript. Also, note that each layer needs its own initial hidden state, .

Now that we've learned about the various architectures of RNN, let's take a look at another aspect of RNN in natural language processing.