Summary

In this chapter, we discussed RNNs. This type of neural network was developed in order to solve problems related to data in sequences. This means that a single instance does not contain all the relevant information since this depends on information from the previous instances.

There are several applications that fit this type of description. For example, a specific portion of text (or speech) may not mean much without the context of the rest of the text. However, even though NLP has been explored the most with RNNs, there are other applications where the context of the text is important, such as forecasting, video processing, or music-related problems.

An RNN works in a very clever way; the network not only outputs a result but also one or more values that are often referred to as memory. This memory value is used as input for future predictions.

When working with data problems that deal with very large sequences, traditional RNNs present a problem called the vanishing...