January 2019
Intermediate to advanced
390 pages
9h 16m
English
Gated recurrent unit (GRU) is another modification of RNN. It has a simplified architecture compared to LSTM and overcomes the vanishing gradient problem. It takes only two inputs, the input xt at time t and memory ht-1 from time t-1. There are only two gates, Update Gate and Reset Gate, shown in the following diagram:

The update gate controls how much previous memory to keep, and the reset gate determines how to combine the new input with previous memory. We can define the complete GRU cell by the following four equations:
Both GRU and LSTM give a comparable performance, but ...
Read now
Unlock full access