The GRU is a variation of the LSTM architecture. They have one less gate (filter), and the connections are implemented differently. In this variant, the forget gate and the input gate are combined into one update gate. Besides this, the cell state and the latent state are combined. The resulting model is simpler than standard LSTM models, and as a result, it is gaining more popularity:

The updated gate determines how much information should remain from the last state and how much should be taken from the previous layer:

The reset gate works like the forget gate:

The tanh layer builds a vector of new candidate values, , that can be added to the state of the cell. The values of the reset gate are applied to the values of the previous state:

A new state is obtained based on a combination of previous state values and new candidate values. The update gate values control in what proportion state values should be used: