Apache Spark Deep Learning Cookbook

How to do it...

At every node/neuron of a recurrent network, a series of matrix multiplication steps are carried out. The input vector/matrix is multiplied by a weight vector/matrix first, a bias is added to this term, and this is finally passed through an activation function to produce the output (just as in the case of feedforward networks):

1. The following diagram shows an intuitive and mathematical way of visualizing RNNs in the form of a computational graph:

2. At the first time step (which is t=0), h₀ is calculated using the first formula on the right-hand side of the preceding diagram. Since h^-1 does not exist, the middle term becomes ...