Apache Spark Deep Learning Cookbook

Getting ready

The main components of a Long Short-Term Memory unit are as follows:

• The input gate
• The forget gate
• The update gate

Each of these gates is made up of a sigmoid layer followed by a pointwise multiplication operation. The sigmoid layer outputs numbers between zero and one. These values describe  how much information of each component is allowed to pass through the respective gate. A value of zero means the gate will allow nothing to pass through it, while a value of one means the gate allows all the information to pass through.

The best way to understand LSTM cells is through computational graphs, just like in the case of recurrent neural networks.

LSTMs were originally developed by Sepp Hochreiter and Jurgen Schmidhuber in ...