recurrent neural networks are a widely used class of neural architectures.
They have, however, two shortcomings. First, it is difficult to understand what
exactly they learn. Second, they tend to work poorly on sequences requiring
long-term memorization, despite having this capacity in