Since RNNs are self connected, in some sense it has a lot more depth than just 2, so they can learn very non-linear features (but the weights of the connections

Since RNNs are self connected, in some sense it has a lot more depth than just 2, so they can learn very non-linear features (but the weights of the connections are shared across the layers).